An Open Tool for Estimating the Environmental Footprint of LLMs

Or click an example to test it

Visualisation

Comparative environmental impact of models

The chart below shows the estimated central values for all catalog models under a standardized inference scenario corresponding to 1 hour of active use: 34.6 interactions/hour, 1000 input tokens, 550 output tokens, and one LLM request per use. The hourly pace is derived from an average reading speed of 238.0 words/min (Brysbaert, 2019) and a project convention of 1 token ≈ 0.75 word.

Benchmarks integrated into the chart, all expressed over one hour or rescaled to a comparable order of magnitude: household electricity from Purdue Extension measurements (fluorescent lamp ≈ 9.3 Wh over 1 h; laptop ≈ 32 Wh over 1 h) and a 1500 W electric space heater rescaled here to 4.1 minutes to obtain ≈ 103.5 Wh, close to the order of magnitude of Claude Opus 4.1 in the inference scenario; for carbon, an average gasoline car benchmark derived from the ICCT (2025) factor retained by the project (235 gCO2e/km), here rescaled to 0.17 km to obtain ≈ 40.0 gCO2e, close to the order of magnitude of Claude Opus 4.1 in the inference scenario.

Trade-off

Inference vs. training impact map

This scatter plot compares each catalog model on two axes at once: standardized inference impact over one hour on the horizontal axis and retained training impact on the vertical axis. Point size follows the retained active parameter count, while colors distinguish providers.

Positioning

Inference bubble chart

This explicit bubble chart positions each model by effective active parameters and by its retained inference impact. Bubble size reflects the retained context window, while colors distinguish providers.

Uncertainty

Inference uncertainty span by model

This chart makes the project’s retained inference range explicit for each model by showing the low, central, and high values under the standardized one-hour scenario.

Positioning

Inference model landscape

This landscape view clusters the catalog models from the characteristics retained by the project for inference screening: active and effective parameters, context window, serving mode, modality support, architecture notes, and central energy and carbon outputs. Nearby points indicate models with similar retained screening profiles, not a simple one-metric ranking.

Proxy

Inference screening factor heatmap

This heatmap exposes the central screening factors retained for each market model. It shows the four multiplicative factors used by the project’s prompt proxy and the resulting ratio between effective and raw active parameters.

Positioning

Inference carbon vs. parameter count

This complementary view places models by retained active parameter count on the horizontal axis and by central inference carbon over one hour on the vertical axis, using logarithmic scaling on both axes.

Sensitivity

Country-mix sensitivity

This view compares central inference energy and carbon over one hour, while coloring each model by the retained electricity-mix country used for carbon recalculation. It helps separate model-size effects from country-mix effects.

Timeline

Inference carbon by model release date

This timeline follows the evolution of the project’s central inference CO2e estimate over time for the OpenAI, Claude, Grok, and Mistral families, using the release month of each model as the horizontal axis.

Perspective

Inference CO2 doubling view

This discussion-oriented chart summarizes the central inference screening values for flagship GPT, Claude, and Grok models as a simple doubling-time reading. It should be read as an interpretation of the retained observatory values, not as a provider-side measurement law.

Under the current central screening profile, the flagship inference series suggests a slower increase than training, because standardized usage, active compute, and per-request serving assumptions damp part of the growth that appears in total model scale.

Visualisation

Comparative training impacts of models

The chart below shows the central values retained for all catalog models across two training indicator families: training energy and direct training CO2e. The current screening method combines retained parameter count, a training-token prior, a training-regime prior, architecture features, and a hardware-class proxy. Under these central screening assumptions, frontier models can reach very large training orders of magnitude. Everyday benchmarks are inserted directly into the list to situate those scales, not to imply direct observed equivalence.

Benchmarks integrated into the chart: household electricity for 2,760,139 households over one year of domestic use, i.e. ≈ 6.90 TWh based on an average consumption of 2,500 kWh per household (RTE, 2021 estimate), and full-flight aviation derived from Klöwer et al. (2025) from 577.97 MtCO2 and 27.45 million commercial flights observed in 2023, i.e. ≈ 6,166,824.9 tCO2e for 292,210 full flights. These comparison points are aligned with the current central screening order of magnitude of Claude Opus 4.1 in the training chart, not with a direct provider-side measurement.

Positioning

Training model landscape

This landscape view clusters the catalog models from the characteristics retained by the project for training screening: retained parameter count, training-token prior, training regime, hardware-class proxy, modality support, architecture notes, and central training energy and carbon outputs. Nearby points indicate similar retained screening profiles rather than a direct ranking on one axis.

Positioning

Training screening factor heatmap

This heatmap exposes the central screening factors retained for each market model in the training proxy. It shows the regime, architecture, and hardware factors together with the retained training-token ratio per parameter.

Uncertainty

Training uncertainty span by model

This view shows the low, central, and high direct training CO2e values retained by the project for each market model. It makes explicit how widely the training proxy can vary once the parameter and token exponents and contextual factors are widened.

Positioning

Training carbon vs. parameter count

This complementary view places models by retained parameter count on the horizontal axis and by direct training CO2e on the vertical axis, using logarithmic scaling on both axes.

Timeline

Training carbon by model release date

This timeline follows the evolution of the project’s retained direct training CO2e estimate over time for the OpenAI, Claude, Grok, and Mistral families, using the release month of each model as the horizontal axis.

Perspective

Training CO2 doubling view

This chart compresses the central training screening values of flagship GPT, Claude, and Grok models into a simple doubling-time interpretation. It is meant as a discussion support to make structural acceleration legible, not as a claim of direct industrial telemetry.

The apparent acceleration is stronger for training because the current screening method compounds retained parameter count, token priors, architecture effects, and hardware assumptions. The resulting doubling pace is therefore a transparent scenario reading, not a universal empirical constant.

Models

41 current models tracked by the project

The table below compares the models tracked by the project under the same inference scenario. For each model, the application shows the central values produced by the project’s multi-factor prompt proxy, both per hour of standardized use and per request.
1800B US
Screening proxy
 96.5 Wh 37.2 gCO2e 2.8 Wh 1.1 gCO2e
1800B US
Screening proxy
 96.5 Wh 37.2 gCO2e 2.8 Wh 1.1 gCO2e
95B* US
Screening proxy
 5.9 Wh 2.3 gCO2e 0.17 Wh 0.0657 gCO2e
95B* US
Screening proxy
 5.9 Wh 2.3 gCO2e 0.17 Wh 0.0657 gCO2e
8B* US
Screening proxy
 0.54 Wh 0.21 gCO2e 0.0155 Wh 0.0060 gCO2e
5.1B active / 117B total US
Comparative reference country
 0.40 Wh 0.16 gCO2e 0.0116 Wh 0.0045 gCO2e
3.6B active / 21B total US
Comparative reference country
 0.26 Wh 0.10 gCO2e 0.00740 Wh 0.0029 gCO2e
2000B* US
Screening proxy
 103.5 Wh 39.9 gCO2e 3.0 Wh 1.2 gCO2e
400B* US
Screening proxy
 22.4 Wh 8.6 gCO2e 0.65 Wh 0.25 gCO2e
175B* US
Screening proxy
 10.2 Wh 3.9 gCO2e 0.30 Wh 0.11 gCO2e
175B* US
Screening proxy
 10.2 Wh 3.9 gCO2e 0.30 Wh 0.11 gCO2e
22B* US
Screening proxy
 1.4 Wh 0.55 gCO2e 0.0412 Wh 0.0159 gCO2e
200B* US
Screening proxy
 12.5 Wh 4.8 gCO2e 0.36 Wh 0.14 gCO2e
80B* US
Screening proxy
 5.2 Wh 2.0 gCO2e 0.15 Wh 0.0581 gCO2e
30B* US
Screening proxy
 2.1 Wh 0.79 gCO2e 0.0594 Wh 0.0229 gCO2e
30B* US
Screening proxy
 2.1 Wh 0.79 gCO2e 0.0594 Wh 0.0229 gCO2e
600B* US
Documented region, country retained as reference
 38.0 Wh 14.6 gCO2e 1.1 Wh 0.42 gCO2e
8B US
Comparative reference country
 0.46 Wh 0.18 gCO2e 0.0133 Wh 0.0051 gCO2e
70B US
Comparative reference country
 3.6 Wh 1.4 gCO2e 0.10 Wh 0.0402 gCO2e
405B US
Comparative reference country
 19.1 Wh 7.4 gCO2e 0.55 Wh 0.21 gCO2e
3B FR
Provider-country proxy
 0.19 Wh 0.0069 gCO2e 0.00560 Wh 0.0002 gCO2e
8B FR
Provider-country proxy
 0.49 Wh 0.0208 gCO2e 0.0142 Wh 0.0006 gCO2e
123B FR
Provider-country proxy
 7.2 Wh 0.29 gCO2e 0.21 Wh 0.0083 gCO2e
24B FR
Provider-country proxy
 1.4 Wh 0.0588 gCO2e 0.0413 Wh 0.0017 gCO2e
22B FR
Provider-country proxy
 1.2 Wh 0.0484 gCO2e 0.0347 Wh 0.0014 gCO2e
7B CN
Provider-country proxy
 0.40 Wh 0.22 gCO2e 0.0117 Wh 0.0063 gCO2e
32B CN
Provider-country proxy
 1.7 Wh 0.93 gCO2e 0.0496 Wh 0.0268 gCO2e
72B CN
Provider-country proxy
 3.7 Wh 2.0 gCO2e 0.11 Wh 0.0579 gCO2e
37B active / 671B total CN
Provider-country proxy
 2.8 Wh 1.5 gCO2e 0.0798 Wh 0.0431 gCO2e
37B active / 671B total CN
Provider-country proxy
 2.9 Wh 1.6 gCO2e 0.0836 Wh 0.0451 gCO2e
175B* US
Screening proxy
 9.2 Wh 3.5 gCO2e 0.26 Wh 0.10 gCO2e
440B* active / 1760B* total US
Screening proxy
 26.0 Wh 10.0 gCO2e 0.75 Wh 0.29 gCO2e
100B* US
Screening proxy
 6.0 Wh 2.3 gCO2e 0.17 Wh 0.0669 gCO2e
78.5B* active / 314B* total US
Documented region, country retained as reference
 5.5 Wh 2.1 gCO2e 0.16 Wh 0.0612 gCO2e
115B* active / 270B* total US
Documented region, country retained as reference
 8.3 Wh 3.2 gCO2e 0.24 Wh 0.0920 gCO2e
530B US
Screening proxy
 26.2 Wh 10.1 gCO2e 0.76 Wh 0.29 gCO2e
178B US
Documented region, country retained as reference
 9.3 Wh 3.6 gCO2e 0.27 Wh 0.10 gCO2e
280B GB
Documented region, country retained as reference
 14.3 Wh 2.6 gCO2e 0.41 Wh 0.0744 gCO2e
130B US
Documented region, country retained as reference
 6.4 Wh 2.5 gCO2e 0.18 Wh 0.0710 gCO2e
137B* US
Documented region, country retained as reference
 7.3 Wh 2.8 gCO2e 0.21 Wh 0.0807 gCO2e
175B US
Documented region, country retained as reference
 8.1 Wh 3.1 gCO2e 0.23 Wh 0.0900 gCO2e

`Retained country` is the country actually used to recalculate CO2 via the electricity mix. When the exact country is not published, the project uses an explicit screening proxy rather than presenting a location as certain.

`*` indicates an estimated parameter count rather than a provider-published value.

The market-model comparison now relies on market_multifactor_prompt_proxy_v1: a prompt-energy screening proxy whose main prompt-level calibration anchor comes from Elsworth et al. (2025), then adjusted by active parameters, context window, serving mode, modality support, architecture overhead, and standardized token volume, and interpreted alongside other inference references.

Models

41 current models with estimated training impacts

This table projects the training orders of magnitude of current models from the indicator families actually available in the literature: training energy derived from emissions when the source country is documented in the electricity-mix table, and direct training CO2e. The current screening proxy combines retained parameter count, a training-token prior, a training-regime prior, architecture features, and a hardware-class proxy. Training energy therefore remains a more fragile screening reconstruction than direct carbon.
1800B 5 589.3 GWh 4 995 128 tCO2e
1800B 5 589.3 GWh 4 995 128 tCO2e
95B* 15.6 GWh 13 914 tCO2e
95B* 15.6 GWh 13 914 tCO2e
8B* 110.4 MWh 98.67 tCO2e
5.1B active / 117B total 21.6 GWh 19 269 tCO2e
3.6B active / 21B total 694.6 MWh 620.8 tCO2e
2000B* 6 900.3 GWh 6 166 825 tCO2e
400B* 276.0 GWh 246 673 tCO2e
175B* 52.8 GWh 47 215 tCO2e
175B* 52.8 GWh 47 215 tCO2e
22B* 834.9 MWh 746.2 tCO2e
200B* 69.0 GWh 61 668 tCO2e
80B* 11.0 GWh 9 867 tCO2e
30B* 1.6 GWh 1 388 tCO2e
30B* 1.6 GWh 1 388 tCO2e
600B* 621.0 GWh 555 014 tCO2e
8B 112.0 MWh 100.1 tCO2e
70B 8.6 GWh 7 664 tCO2e
405B 287.1 GWh 256 543 tCO2e
3B 15.0 MWh 13.41 tCO2e
8B 106.7 MWh 95.33 tCO2e
123B 26.1 GWh 23 324 tCO2e
24B 1.1 GWh 986.7 tCO2e
22B 806.7 MWh 721.0 tCO2e
7B 85.8 MWh 76.64 tCO2e
32B 1.8 GWh 1 602 tCO2e
72B 9.1 GWh 8 108 tCO2e
37B active / 671B total 675.4 GWh 603 599 tCO2e
37B active / 671B total 675.4 GWh 603 599 tCO2e
175B* 53.6 GWh 47 899 tCO2e
440B* active / 1760B* total 819.8 GWh 732 619 tCO2e
100B* 17.3 GWh 15 417 tCO2e
78.5B* active / 314B* total 153.1 GWh 136 805 tCO2e
115B* active / 270B* total 113.2 GWh 101 151 tCO2e
530B 795.0 GWh 710 525 tCO2e
178B 77.9 GWh 69 600 tCO2e
280B 126.0 GWh 112 612 tCO2e
130B 35.1 GWh 31 370 tCO2e
137B* 30.8 GWh 27 550 tCO2e
175B 61.3 GWh 54 742 tCO2e

`*` indicates an estimated parameter count rather than a provider-published value.

ImpactLLM is designed as a transparent screening tool, not as a black-box score. The current release starts from source-linked inference anchors, then exposes a bounded multi-factor proxy rather than a hidden single-number score.

1. Source-linked literature anchors.

The application-level estimator starts from published inference indicators linked to an explicit source, model, geography, and system boundary. In the current market-model release, the predictive core uses Elsworth et al. (2025) as the main prompt-level calibration anchor, with a median prompt energy of 0.24 Wh/prompt for Gemini Apps, and is interpreted alongside other inference references such as the ML.ENERGY Benchmark, Ren et al. (2024), and Li et al. (2025).

2. A multi-factor effective-parameter proxy.

When direct telemetry is unavailable for a target model, ImpactLLM does not rely on a raw parameter multiple alone. It builds an effective active-parameter profile from the retained model characteristics: active parameters, context window, serving mode (open, hybrid, closed), modality support, and architecture notes such as MoE or reasoning-oriented overheads.

3. Token volume remains explicit.

The current proxy adjusts the anchor with a weighted prompt-compute volume defined from input and output tokens. Output generation is weighted more heavily than input processing, so output-heavy scenarios and repeated LLM calls raise the estimate materially.

The current prompt-level branch is a screening proxy, not an audited benchmark. For this reason, the application returns a bounded low-central-high result rather than one falsely precise deterministic value.

4. Carbon derived from context.

Carbon is not copied mechanically from the source paper. It is recalculated from the retained energy estimate using the electricity mix associated with the selected country context.

5. A research-oriented estimator.

The result is an auditable estimate intended for comparison, software design, and methodological discussion. It is useful precisely because the assumptions, factors, and retained sources remain visible and inspectable.

Technical paper

Preview of the first page of Transparent Screening for LLM Inference and Training Impacts Technical paper PDF

Pachot, A., & Petit, T. (2026, March 14). Transparent Screening for LLM Inference and Training Impacts. /impact-llm/downloads/ImpactLLM_paper.pdf

We work on responsible AI with a focus on methodological rigor, traceability, and real-world decision support. Our work combines scientific research, product design, and operational deployment to make AI systems more transparent, more accountable, and more useful in practice.

How to cite ImpactLLM

Pachot, A., & Petit, T. (2026, March 14). Transparent Screening for LLM Inference and Training Impacts.

BibTeX

@misc{impactllm_screening_2026,
  title = {Transparent Screening for LLM Inference and Training Impacts},
  author = {Pachot, Arnault and Petit, Thierry},
  year = {2026},
  month = mar,
  note = {Working paper},
  url = {https://dev.emotia.com/impact-llm/downloads/ImpactLLM_paper.pdf}
}

Download technical paper PDF | Download technical paper BibTeX

GitHub repository

The project repository is available on GitHub: https://github.com/apachot/ImpactLLM.

License

This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with this program. If not, see https://www.gnu.org/licenses/.

Arnault Pachot

Arnault Pachot is a researcher and entrepreneur, founder of OpenStudio and now founder of Emotia. He works on responsible digital transformation, Green IT, and decision-oriented AI systems. He co-authored the Dunod book Intelligence artificielle et environnement : alliance ou nuisance ?, dedicated to practical pathways for environmentally responsible AI.

LinkedIn: Arnault Pachot

Google Scholar: Arnault Pachot

Cover of the Dunod book Intelligence artificielle et environnement : alliance ou nuisance ? Arnault Pachot's Dunod book

Thierry Petit

Thierry Petit is a senior AI researcher and scientific leader with more than twenty years of academic and R&D experience in Europe and the United States. His work spans trustworthy AI, simulation, optimization, and decision-grade platforms. At Emotia and Pollitics, he leads the scientific direction of systems designed to remain both operationally useful and methodologically robust.

LinkedIn: Thierry Petit

Google Scholar: Thierry Petit

Selected references on AI and the environment

Sources

This annex brings together the quantified reference material used in the interface, along with everyday comparison benchmarks and country factors used for carbon and water recalculation.

Inference reference set

Ref. Data type LLM model Parameters Country Value Citation
[37]Energy consumption per prompt (Gemini Apps median prompt)Gemini Apps (version non spécifiée)180BNon spécifié0.24 Wh/promptElsworth, C., Huang, K., Patterson, D., Schneider, I., & others (2025). Measuring the Environmental Impact of Delivering AI at Google Scale. arXiv preprint arXiv:2508.15734. https://arxiv.org/abs/2508.15734
Table 2, p. 7
[38]Emissions per prompt (Gemini Apps median prompt)Gemini Apps (version non spécifiée)180BNon spécifié0.03 gCO2e/promptElsworth, C., Huang, K., Patterson, D., Schneider, I., & others (2025). Measuring the Environmental Impact of Delivering AI at Google Scale. arXiv preprint arXiv:2508.15734. https://arxiv.org/abs/2508.15734
Table 2, p. 7
[40]Energy consumption to generate one page (Llama-3-70B, 500-word page)Llama 3 70B70BÉtats-Unis0.0195 kWh/pageRen, S., Tomlinson, B., Black, R. W., & Torrance, A. W. (2024). Reconciling the Contrasting Narratives on the Environmental Impact of Large Language Models. Scientific Reports, 14, 28180. https://www.nature.com/articles/s41598-024-76682-6
Results section, Llama-3-70B paragraphs
[41]Emissions to generate one page (Llama-3-70B, 500-word page)Llama 3 70B70BÉtats-Unis15 gCO2/pageRen, S., Tomlinson, B., Black, R. W., & Torrance, A. W. (2024). Reconciling the Contrasting Narratives on the Environmental Impact of Large Language Models. Scientific Reports, 14, 28180. https://www.nature.com/articles/s41598-024-76682-6
Results section, Llama-3-70B paragraphs
[43]Energy consumption to generate one page (Gemma-2B-it, 500-word page)Gemma-2B-it2BÉtats-Unis0.00024 kWh/pageRen, S., Tomlinson, B., Black, R. W., & Torrance, A. W. (2024). Reconciling the Contrasting Narratives on the Environmental Impact of Large Language Models. Scientific Reports, 14, 28180. https://www.nature.com/articles/s41598-024-76682-6
Results section, Gemma-2B-it paragraphs
[44]Emissions to generate one page (Gemma-2B-it, 500-word page)Gemma-2B-it2BÉtats-Unis0.18 gCO2/pageRen, S., Tomlinson, B., Black, R. W., & Torrance, A. W. (2024). Reconciling the Contrasting Narratives on the Environmental Impact of Large Language Models. Scientific Reports, 14, 28180. https://www.nature.com/articles/s41598-024-76682-6
Results section, Gemma-2B-it paragraphs

Training reference set

Ref. Data type LLM model Parameters Country Value Citation
[1]Greenhouse gas emissions from training (Transformer (big))Transformer (big)213MÉtats-Unis192 lb CO2eStrubell, E., Ganesh, A., & McCallum, A. (2019). Energy and Policy Considerations for Deep Learning in NLP. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 3645--3650. https://aclanthology.org/P19-1355/
Table 1, p. 1
[2]Greenhouse gas emissions from training (BLOOM 176B)BLOOM 176B176BFrance24.7 tCO2eLuccioni, A. S., Viguier, S., & Ligozat, A. L. (2023). Estimating the Carbon Footprint of BLOOM. Journal of Machine Learning Research, 24(253), 1--15. https://www.jmlr.org/papers/v24/23-0069.html
Abstract, p. 1
[3]Greenhouse gas emissions from training (BLOOM 176B)BLOOM 176B176BFrance50.5 tCO2eLuccioni, A. S., Viguier, S., & Ligozat, A. L. (2023). Estimating the Carbon Footprint of BLOOM. Journal of Machine Learning Research, 24(253), 1--15. https://www.jmlr.org/papers/v24/23-0069.html
Abstract, p. 1; Table 3, p. 7
[4]Greenhouse gas emissions from training (Llama 3.1 8B)Llama 3.1 8B8BNon spécifié420 tCO2eMeta (2024). Llama 3.1 Model Card. https://huggingface.co/meta-llama/Llama-3.1-405B
Section 'Hardware and Software', table 'Training Location-Based Greenhouse Gas Emissions'
[5]Greenhouse gas emissions from training (Llama 3.1 70B)Llama 3.1 70B70BNon spécifié2040 tCO2eMeta (2024). Llama 3.1 Model Card. https://huggingface.co/meta-llama/Llama-3.1-405B
Section 'Hardware and Software', table 'Training Location-Based Greenhouse Gas Emissions'
[6]Greenhouse gas emissions from training (Llama 3.1 405B)Llama 3.1 405B405BNon spécifié8930 tCO2eMeta (2024). Llama 3.1 Model Card. https://huggingface.co/meta-llama/Llama-3.1-405B
Section 'Hardware and Software', table 'Training Location-Based Greenhouse Gas Emissions'
[7]Compute time used for training (Llama 3.1 405B)Llama 3.1 405B405BNon spécifié30.84 million GPU-hoursMeta (2024). Llama 3.1 Model Card. https://huggingface.co/meta-llama/Llama-3.1-405B
Section 'Hardware and Software', cumulative compute table
[8]Compute time used for training (Llama 3.1 8B)Llama 3.1 8B8BNon spécifié1.46 million GPU-hoursMeta (2024). Llama 3.1 Model Card. https://huggingface.co/meta-llama/Llama-3.1-405B
Section 'Hardware and Software', cumulative compute table
[9]Compute time used for training (Llama 3.1 70B)Llama 3.1 70B70BNon spécifié7.0 million GPU-hoursMeta (2024). Llama 3.1 Model Card. https://huggingface.co/meta-llama/Llama-3.1-405B
Section 'Hardware and Software', cumulative compute table
[10]Training token volume (Llama 3.1 8B)Llama 3.1 8B8BNon spécifié15 trillion tokensMeta (2024). Llama 3.1 Model Card. https://huggingface.co/meta-llama/Llama-3.1-405B
Section 'Training Data'
[11]Training token volume (Llama 3.1 70B)Llama 3.1 70B70BNon spécifié15 trillion tokensMeta (2024). Llama 3.1 Model Card. https://huggingface.co/meta-llama/Llama-3.1-405B
Section 'Training Data'
[12]Training token volume (Llama 3.1 405B)Llama 3.1 405B405BNon spécifié15 trillion tokensMeta (2024). Llama 3.1 Model Card. https://huggingface.co/meta-llama/Llama-3.1-405B
Section 'Training Data'
[13]Emissions across the model creation lifecycle (OLMo 20M)OLMo 20M20MÉtats-Unis0.3 tCO2eMorrison, J., Na, C., Fernandez, J., Dettmers, T., & others (2025). Holistically Evaluating the Environmental Impact of Creating Language Models. arXiv preprint arXiv:2503.05804. https://arxiv.org/abs/2503.05804
Table 2, p. 6; cluster locations in Section 3.1, p. 4
[14]Emissions across the model creation lifecycle (OLMo 60M)OLMo 60M60MÉtats-Unis0.4 tCO2eMorrison, J., Na, C., Fernandez, J., Dettmers, T., & others (2025). Holistically Evaluating the Environmental Impact of Creating Language Models. arXiv preprint arXiv:2503.05804. https://arxiv.org/abs/2503.05804
Table 2, p. 6; cluster locations in Section 3.1, p. 4
[15]Emissions across the model creation lifecycle (OLMo 150M)OLMo 150M150MÉtats-Unis1 tCO2eMorrison, J., Na, C., Fernandez, J., Dettmers, T., & others (2025). Holistically Evaluating the Environmental Impact of Creating Language Models. arXiv preprint arXiv:2503.05804. https://arxiv.org/abs/2503.05804
Table 2, p. 6; cluster locations in Section 3.1, p. 4
[16]Emissions across the model creation lifecycle (OLMo 300M)OLMo 300M300MÉtats-Unis2 tCO2eMorrison, J., Na, C., Fernandez, J., Dettmers, T., & others (2025). Holistically Evaluating the Environmental Impact of Creating Language Models. arXiv preprint arXiv:2503.05804. https://arxiv.org/abs/2503.05804
Table 2, p. 6; cluster locations in Section 3.1, p. 4
[17]Emissions across the model creation lifecycle (OLMo 700M)OLMo 700M700MÉtats-Unis3 tCO2eMorrison, J., Na, C., Fernandez, J., Dettmers, T., & others (2025). Holistically Evaluating the Environmental Impact of Creating Language Models. arXiv preprint arXiv:2503.05804. https://arxiv.org/abs/2503.05804
Table 2, p. 6; cluster locations in Section 3.1, p. 4
[18]Emissions across the model creation lifecycle (OLMo 7B)OLMo 7B7BÉtats-Unis22 tCO2eMorrison, J., Na, C., Fernandez, J., Dettmers, T., & others (2025). Holistically Evaluating the Environmental Impact of Creating Language Models. arXiv preprint arXiv:2503.05804. https://arxiv.org/abs/2503.05804
Table 2, p. 6; cluster locations in Section 3.1, p. 4
[19]Emissions across the model creation lifecycle (OLMo 1B (3T))OLMo 1B (3T)1BÉtats-Unis10 tCO2eMorrison, J., Na, C., Fernandez, J., Dettmers, T., & others (2025). Holistically Evaluating the Environmental Impact of Creating Language Models. arXiv preprint arXiv:2503.05804. https://arxiv.org/abs/2503.05804
Table 2, p. 6; cluster locations in Section 3.1, p. 4
[20]Emissions across the model creation lifecycle (OLMo 7B (Twin))OLMo 7B (Twin)7BÉtats-Unis70 tCO2eMorrison, J., Na, C., Fernandez, J., Dettmers, T., & others (2025). Holistically Evaluating the Environmental Impact of Creating Language Models. arXiv preprint arXiv:2503.05804. https://arxiv.org/abs/2503.05804
Table 2, p. 6; cluster locations in Section 3.1, p. 4
[21]Emissions across the model creation lifecycle (OLMo (04|07)24 7B)OLMo (04|07)24 7B7BÉtats-Unis32 tCO2eMorrison, J., Na, C., Fernandez, J., Dettmers, T., & others (2025). Holistically Evaluating the Environmental Impact of Creating Language Models. arXiv preprint arXiv:2503.05804. https://arxiv.org/abs/2503.05804
Table 2, p. 6; cluster locations in Section 3.1, p. 4
[22]Emissions across the model creation lifecycle (OLMo 2 7B)OLMo 2 7B7BÉtats-Unis52 tCO2eMorrison, J., Na, C., Fernandez, J., Dettmers, T., & others (2025). Holistically Evaluating the Environmental Impact of Creating Language Models. arXiv preprint arXiv:2503.05804. https://arxiv.org/abs/2503.05804
Table 2, p. 6; cluster locations in Section 3.1, p. 4
[23]Emissions across the model creation lifecycle (OLMo 2 13B)OLMo 2 13B13BÉtats-Unis101 tCO2eMorrison, J., Na, C., Fernandez, J., Dettmers, T., & others (2025). Holistically Evaluating the Environmental Impact of Creating Language Models. arXiv preprint arXiv:2503.05804. https://arxiv.org/abs/2503.05804
Table 2, p. 6; cluster locations in Section 3.1, p. 4
[24]Emissions across the model creation lifecycle (OLMoE 0924)OLMoE 09241B active / 7B totalÉtats-Unis18 tCO2eMorrison, J., Na, C., Fernandez, J., Dettmers, T., & others (2025). Holistically Evaluating the Environmental Impact of Creating Language Models. arXiv preprint arXiv:2503.05804. https://arxiv.org/abs/2503.05804
Table 2, p. 6; cluster locations in Section 3.1, p. 4

Real-world comparison benchmarks

No. Domain Indicator Value Reference
1 Energy Fluorescent lamp for 1 hour 9,3 Wh Carroll, N. J., & Kruse, J. (n.d.). Energy investigators 2: Facilitator’s guide. Purdue Extension. https://www.extension.purdue.edu/extmedia/4H/4-H-1015-W.pdf
Purdue Extension guide used for household-use examples; extraction still to be normalized more precisely.
2 Energy Laptop for 1 hour 32 Wh Carroll, N. J., & Kruse, J. (n.d.). Energy investigators 2: Facilitator’s guide. Purdue Extension. https://www.extension.purdue.edu/extmedia/4H/4-H-1015-W.pdf
Purdue Extension guide used for household-use examples; extraction still to be normalized more precisely.
3 Energy Electric space heater for 1 hour 1.5 kWh Project calculation convention.
Nominal power assumption fixed at 1,500 W, i.e. 1.5 kWh for 1 hour.
4 Energy 10,000 French households over one year of domestic use 25 GWh RTE. (2022, February 25). Bilan électrique 2021 - Une production d’électricité assurée à plus de 92% par des sources n’émettant pas de gaz à effet de serre. https://www.rte-france.com/actualites/bilan-electrique-2021
Project convention based on an average consumption of 2,500 kWh/year per household; comparison value made explicit in the training-chart note.
5 Carbon Average internal-combustion car 235 gCO2e/km International Council on Clean Transportation. (2025). Life-cycle greenhouse gas emissions from passenger cars in the European Union: A 2025 update and key factors to consider. https://theicct.org/publication/electric-cars-life-cycle-analysis-emissions-europe-jul25/
Key findings ; Figure 1 ; gasoline ICEV running on the average blend of fossil gasoline and ethanol estimated at 235 gCO2e/km.
6 Carbon Average gasoline car for 0.17 km 40.0 gCO2e International Council on Clean Transportation. (2025). Life-cycle greenhouse gas emissions from passenger cars in the European Union: A 2025 update and key factors to consider. https://theicct.org/publication/electric-cars-life-cycle-analysis-emissions-europe-jul25/
Project derivation from the ICCT benchmark of 235 gCO2e/km for an average gasoline car, rescaled here to 0.17 km to align with the order of magnitude of Claude Opus 4.1 in the inference comparison.
7 Carbon Average full commercial flight (derived value) ≈ 21.1 tCO2 per flight Gössling, S., Klöwer, M., Leitão, J. C., Hirsch, S., Brockhagen, D., & Humpe, A. (2026). Large carbon dioxide emissions avoidance potential in improved commercial air transport efficiency. Communications Earth & Environment, 7, 13. https://www.nature.com/articles/s43247-025-03069-4
Results, “Emissions and efficiency”: 27,451,887 flights in 2023 causing 577,968,750 tCO2 emissions; the site then derives an average per flight.

Central screening factors retained for market models

This table documents the central values retained by the project for the multi-factor prompt proxy of each catalog model: raw active parameters, context window, serving mode, modality support, the resulting central factors F_ctx, F_srv, F_mod, F_arch, and the resulting central effective active-parameter proxy P_eff,c. These are project screening factors, not provider-published measurements.

Model Provider Active parameters Context window Serving mode Vision F_ctx F_srv F_mod F_arch P_eff,c
ai21 178B [1] 2,048 [2] closed [4] no [3] 1.000 1.140 1.000 1.000 202.920B
alibaba 32B [9] 131,072 [10] open [12] no [11] 1.070 1.000 1.000 1.000 34.240B
alibaba 72B [18] 131,072 [19] open [21] no [20] 1.070 1.000 1.000 1.000 77.040B
alibaba 7B [24] 131,072 [25] open [27] no [26] 1.070 1.000 1.000 1.000 7.490B
anthropic 100B* [30] 200,000 [31] closed [33] yes [32] 1.091 1.140 1.030 1.000 128.145B
anthropic 22B* [37] 200,000 [38] closed [40] yes [39] 1.091 1.140 1.030 1.000 28.192B
anthropic 175B* [43] 200,000 [38] closed [40] yes [39] 1.091 1.140 1.030 1.000 224.253B
anthropic 175B* [45] 200,000 [38] closed [40] yes [39] 1.091 1.140 1.030 1.000 224.253B
anthropic 2000B* [47] 200,000 [38] closed [40] yes [39] 1.091 1.140 1.030 1.000 2562.897B
anthropic 400B* [49] 200,000 [38] closed [40] yes [39] 1.091 1.140 1.030 1.000 512.579B
deepmind 280B [51] 4,096 [52] closed [54] no [53] 1.000 1.140 1.000 1.000 319.200B
deepseek 37B active / 671B total [59] 128,000 [60] hybrid [62] no [61] 1.069 1.070 1.000 1.401 59.289B
deepseek 37B active / 671B total [65] 128,000 [66] hybrid [68] no [67] 1.069 1.070 1.000 1.334 56.466B
google 30B* [71] 1,048,576 [72] closed [74] yes [73] 1.175 1.140 1.030 1.000 41.391B
google 80B* [77] 1,048,576 [72] closed [74] yes [73] 1.175 1.140 1.030 1.000 110.375B
google 30B* [79] 1,048,576 [72] closed [74] yes [73] 1.175 1.140 1.030 1.000 41.391B
google 200B* [81] 1,048,576 [72] closed [74] yes [73] 1.175 1.140 1.030 1.000 275.937B
google 130B [83] 2,048 [84] research [86] no [85] 1.000 1.050 1.000 1.000 136.500B
google 137B* [91] 2,048 [92] closed [94] no [93] 1.000 1.140 1.000 1.000 156.180B
meta 405B [99] 131,072 [100] open [102] no [101] 1.070 1.000 1.000 1.000 433.350B
meta 70B [99] 131,072 [105] open [107] no [106] 1.070 1.000 1.000 1.000 74.900B
meta 8B [99] 131,072 [109] open [111] no [110] 1.070 1.000 1.000 1.000 8.560B
meta 175B [113] 2,048 [114] open [116] no [115] 1.000 1.000 1.000 1.000 175.000B
microsoft 530B [121] 2,048 [122] closed [124] no [123] 1.000 1.140 1.000 1.000 604.200B
mistral 22B [129] 32,000 [130] hybrid [132] no [131] 1.000 1.070 1.000 1.000 23.540B
mistral 3B [129] 128,000 [135] hybrid [137] no [136] 1.069 1.070 1.000 1.000 3.431B
mistral 8B [129] 128,000 [139] hybrid [141] no [140] 1.069 1.070 1.000 1.000 9.149B
mistral 123B [143] 128,000 [144] closed [146] yes [145] 1.069 1.140 1.030 1.000 154.364B
mistral 24B [129] 128,000 [149] hybrid [151] yes [150] 1.069 1.070 1.030 1.000 28.270B
openai 175B* [153] 4,096 [154] closed [156] no [155] 1.000 1.140 1.000 1.000 199.500B
openai 440B* active / 1760B* total [161] 8,192 [162] closed [164] yes [163] 1.000 1.140 1.030 1.160 599.312B
openai 8B* [170] 128,000 [171] closed [173] yes [172] 1.069 1.140 1.030 1.000 10.040B
openai 95B* [178] 400,000 [179] closed [181] yes [180] 1.126 1.140 1.030 1.000 125.642B
openai 95B* [184] 400,000 [185] closed [187] yes [186] 1.126 1.140 1.030 1.000 125.642B
openai 1800B [190] 400,000 [191] closed [193] yes [192] 1.126 1.140 1.030 1.000 2380.582B
openai 1800B [190] 400,000 [196] closed [198] yes [197] 1.126 1.140 1.030 1.000 2380.582B
openai 5.1B active / 117B total [200] 131,072 [201] open [203] no [202] 1.070 1.000 1.000 1.362 7.430B
openai 3.6B active / 21B total [206] 131,072 [207] open [209] no [208] 1.070 1.000 1.000 1.204 4.636B
xai 78.5B* active / 314B* total [212] 200,000 [213] closed [215] yes [214] 1.091 1.140 1.030 1.160 116.689B
xai 115B* active / 270B* total [219] 2,000,000 [213] closed [215] yes [214] 1.208 1.140 1.030 1.099 179.130B
xai 600B* [222] 2,000,000 [223] closed [225] yes [224] 1.208 1.140 1.030 1.050 893.321B

Central training screening factors retained for market models

This table documents the central values retained by the project for the multi-factor training proxy of each catalog model: retained training parameter count, training-token prior, training regime, multimodal training flag, hardware-class proxy, and the resulting central factors F_reg, F_arch-tr, and F_hw. These are project screening factors, not provider-published measurements.

Model Provider Retained parameters Training tokens Training regime Multimodal Hardware class F_reg F_arch F_hw
ai21 178B [5] 5.00T [6] pretraining [7] no [8] standard_gpu_cluster 1.0000 1.0000 1.0500
alibaba 32B [13] 0.64T [14] pretraining [15] no [16] standard_gpu_cluster [17] 1.0000 1.0000 1.0500
alibaba 72B [22] 1.44T [14] pretraining [15] no [23] standard_gpu_cluster [17] 1.0000 1.0000 1.0500
alibaba 7B [28] 0.14T [14] pretraining [15] no [29] standard_gpu_cluster [17] 1.0000 1.0000 1.0500
anthropic 100B [34] 2.00T [14] pretraining [35] yes [36] modern_hyperscale_gpu 1.0000 1.1500 0.9000
anthropic 22B [41] 0.44T [14] pretraining [15] yes [42] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
anthropic 175B [44] 3.50T [14] pretraining [15] yes [42] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
anthropic 175B [46] 3.50T [14] pretraining [15] yes [42] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
anthropic 2000B [48] 40.00T [14] pretraining [15] yes [42] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
anthropic 400B [50] 8.00T [14] pretraining [15] yes [42] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
deepmind 280B [55] 6.00T [56] pretraining [57] no [58] modern_hyperscale_gpu 1.0000 1.0000 0.9000
deepseek 671B [63] 13.42T [14] pretraining [15] no [64] mixed_gpu_cluster [17] 1.0000 0.9000 1.0000
deepseek 671B [69] 13.42T [14] pretraining [15] no [70] mixed_gpu_cluster [17] 1.0000 0.9000 1.0000
google 30B [75] 0.60T [14] pretraining [15] yes [76] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
google 80B [78] 1.60T [14] pretraining [15] yes [76] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
google 30B [80] 0.60T [14] pretraining [15] yes [76] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
google 200B [82] 4.00T [14] pretraining [15] yes [76] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
google 130B [87] 4.00T [88] pretraining [89] no [90] modern_hyperscale_gpu 1.0000 0.9000 0.9000
google 137B [95] 3.00T [96] pretraining [97] no [98] modern_hyperscale_gpu 1.0000 1.0000 0.9000
meta 405B [103] 8.10T [14] pretraining [15] no [104] standard_gpu_cluster [17] 1.0000 1.0000 1.0500
meta 70B [103] 1.40T [14] pretraining [15] no [108] standard_gpu_cluster [17] 1.0000 1.0000 1.0500
meta 8B [103] 0.16T [14] pretraining [15] no [112] standard_gpu_cluster [17] 1.0000 1.0000 1.0500
meta 175B [117] 4.00T [118] pretraining [119] no [120] standard_gpu_cluster 1.0000 1.0000 1.0500
microsoft 530B [125] 20.00T [126] pretraining [127] no [128] modern_hyperscale_gpu 1.0000 1.0000 0.9000
mistral 22B [133] 0.44T [14] pretraining [15] no [134] mixed_gpu_cluster [17] 1.0000 1.0000 1.0000
mistral 3B [133] 0.06T [14] pretraining [15] no [138] mixed_gpu_cluster [17] 1.0000 1.0000 1.0000
mistral 8B [133] 0.16T [14] pretraining [15] no [142] mixed_gpu_cluster [17] 1.0000 1.0000 1.0000
mistral 123B [147] 2.46T [14] pretraining [15] yes [148] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
mistral 24B [133] 0.48T [14] pretraining [15] yes [152] mixed_gpu_cluster [17] 1.0000 1.1500 1.0000
openai 175B [157] 3.50T [14] pretraining [158] no [159] standard_gpu_cluster [160] 1.0000 1.0000 1.0500
openai 1760B [165] 6.00T [166] pretraining [167] yes [168] modern_hyperscale_gpu [169] 1.0000 1.0350 0.9000
openai 8B [174] 0.16T [14] pretraining [175] yes [176] modern_hyperscale_gpu [177] 1.0000 1.1500 0.9000
openai 95B [182] 1.90T [14] pretraining [15] yes [183] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
openai 95B [188] 1.90T [14] pretraining [15] yes [189] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
openai 1800B [194] 36.00T [14] pretraining [15] yes [195] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
openai 1800B [194] 36.00T [14] pretraining [15] yes [199] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000
openai 117B [204] 2.34T [14] pretraining [15] no [205] standard_gpu_cluster [17] 1.0000 0.9000 1.0500
openai 21B [210] 0.42T [14] pretraining [15] no [211] standard_gpu_cluster [17] 1.0000 0.9000 1.0500
xai 314B [216] 6.28T [14] pretraining [217] yes [218] modern_hyperscale_gpu 1.0000 1.0350 0.9000
xai 270B [220] 5.40T [14] pretraining [221] yes [218] modern_hyperscale_gpu 1.0000 1.0350 0.9000
xai 600B [226] 12.00T [14] pretraining [15] yes [227] modern_hyperscale_gpu [17] 1.0000 1.1500 0.9000

Numbered source list for retained screening characteristics

The numbered references used in the retained inference and training screening-characteristic tables are listed below.

  1. [1] Active parameters. AI21 Jurassic-1 announcement
  2. [2] Context window. AI21 admissions
  3. [3] Vision. AI21 blog
  4. [4] Serving mode. AI21 admissions
  5. [5] Training parameters. AI21 Jurassic-1 announcement
  6. [6] Training tokens. AI21 Jurassic-1 announcement
  7. [7] Training regime. Jurassic-1 announcement
  8. [8] Training modality. Jurassic-1 platform blog describing text API
  9. [9] Active parameters. Qwen2.5 32B model card
  10. [10] Context window. Qwen2.5 32B model card
  11. [11] Vision. Qwen2.5 32B model card
  12. [12] Serving mode. Qwen2.5 32B model card
  13. [13] Training parameters. Qwen2.5 32B model card
  14. [14] Training tokens. Project screening prior (20 training tokens per retained parameter)
  15. [15] Training regime. Project screening prior (foundation-model pretraining default)
  16. [16] Training modality. Qwen2.5 32B model card
  17. [17] Training hardware. Project screening prior from market status and serving mode
  18. [18] Active parameters. Qwen2.5 72B model card
  19. [19] Context window. Qwen2.5 72B model card
  20. [20] Vision. Qwen2.5 72B model card
  21. [21] Serving mode. Qwen2.5 72B model card
  22. [22] Training parameters. Qwen2.5 72B model card
  23. [23] Training modality. Qwen2.5 72B model card
  24. [24] Active parameters. Qwen2.5 7B model card
  25. [25] Context window. Qwen2.5 7B model card
  26. [26] Vision. Qwen2.5 7B model card
  27. [27] Serving mode. Qwen2.5 7B model card
  28. [28] Training parameters. Qwen2.5 7B model card
  29. [29] Training modality. Qwen2.5 7B model card
  30. [30] Active parameters. Third-party estimate (Claude 2)
  31. [31] Context window. Anthropic Claude overview
  32. [32] Vision. Anthropic Claude overview
  33. [33] Serving mode. Anthropic Claude overview
  34. [34] Training parameters. Third-party estimate (Claude 2)
  35. [35] Training regime. Claude 2 announcement | Anthropic
  36. [36] Training modality. https://docs.anthropic.com/en/docs/about-claude/models/overview
  37. [37] Active parameters. Artificial Analysis size taxonomy for proprietary/open-weight classes; small class midpoint (4B–40B) used as screening estimate
  38. [38] Context window. Models overview - Claude API Docs
  39. [39] Vision. Models overview - Claude API Docs
  40. [40] Serving mode. Models overview - Claude API Docs
  41. [41] Training parameters. Artificial Analysis size taxonomy for proprietary/open-weight classes; small class midpoint (4B–40B) used as screening estimate
  42. [42] Training modality. Models overview - Claude API Docs
  43. [43] Active parameters. Claude 3.7 Sonnet external estimate reused as a family proxy for Claude 3.5 Sonnet
  44. [44] Training parameters. Claude 3.7 Sonnet external estimate reused as a family proxy for Claude 3.5 Sonnet
  45. [45] Active parameters. Alan D. Thompson Models Table estimate for Claude 3.7 Sonnet (175B)
  46. [46] Training parameters. Alan D. Thompson Models Table estimate for Claude 3.7 Sonnet (175B)
  47. [47] Active parameters. Alan D. Thompson Models Table estimate for Claude Opus 4.1 (2T)
  48. [48] Training parameters. Alan D. Thompson Models Table estimate for Claude Opus 4.1 (2T)
  49. [49] Active parameters. Alan D. Thompson Models Table estimate for Claude Sonnet 4.5 (400B) used as family proxy
  50. [50] Training parameters. Alan D. Thompson Models Table estimate for Claude Sonnet 4.5 (400B) used as family proxy
  51. [51] Active parameters. DeepMind Gopher paper
  52. [52] Context window. DeepMind blog
  53. [53] Vision. DeepMind blog
  54. [54] Serving mode. DeepMind blog
  55. [55] Training parameters. DeepMind Gopher paper
  56. [56] Training tokens. DeepMind Gopher release
  57. [57] Training regime. Scaling language models: Gopher
  58. [58] Training modality. DeepMind Gopher paper centered on text modeling
  59. [59] Active parameters. DeepSeek-R1 model card
  60. [60] Context window. deepseek-ai/DeepSeek-R1
  61. [61] Vision. deepseek-ai/DeepSeek-R1
  62. [62] Serving mode. deepseek-ai/DeepSeek-R1
  63. [63] Training parameters. DeepSeek-R1 model card
  64. [64] Training modality. deepseek-ai/DeepSeek-R1
  65. [65] Active parameters. DeepSeek-V3 model card
  66. [66] Context window. deepseek-ai/DeepSeek-V3
  67. [67] Vision. deepseek-ai/DeepSeek-V3
  68. [68] Serving mode. deepseek-ai/DeepSeek-V3
  69. [69] Training parameters. DeepSeek-V3 model card
  70. [70] Training modality. deepseek-ai/DeepSeek-V3
  71. [71] Active parameters. Alan D. Thompson Models Table estimate for Gemini 2.0 Flash exp (30B)
  72. [72] Context window. Gemini models | Gemini API
  73. [73] Vision. Gemini models | Gemini API
  74. [74] Serving mode. Gemini models | Gemini API
  75. [75] Training parameters. Alan D. Thompson Models Table estimate for Gemini 2.0 Flash exp (30B)
  76. [76] Training modality. Gemini models | Gemini API
  77. [77] Active parameters. Alan D. Thompson Models Table estimate for Gemini 2.5 Flash Preview (80B)
  78. [78] Training parameters. Alan D. Thompson Models Table estimate for Gemini 2.5 Flash Preview (80B)
  79. [79] Active parameters. Gemini 2.0 Flash exp estimate (30B) used as family proxy for Flash-Lite
  80. [80] Training parameters. Gemini 2.0 Flash exp estimate (30B) used as family proxy for Flash-Lite
  81. [81] Active parameters. Alan D. Thompson Models Table estimate for Gemini 2.5 Pro (200B)
  82. [82] Training parameters. Alan D. Thompson Models Table estimate for Gemini 2.5 Pro (200B)
  83. [83] Active parameters. Google AI blog
  84. [84] Context window. Google AI blog
  85. [85] Vision. Google AI blog
  86. [86] Serving mode. Google AI blog
  87. [87] Training parameters. Google AI blog
  88. [88] Training tokens. Google AI GLaM blog
  89. [89] Training regime. GLaM release | Google
  90. [90] Training modality. GLaM blog details conditional computation for text
  91. [91] Active parameters. Google presentation
  92. [92] Context window. Google blog
  93. [93] Vision. Google blog
  94. [94] Serving mode. Google blog
  95. [95] Training parameters. Google presentation
  96. [96] Training tokens. Google LaMDA blog
  97. [97] Training regime. LaMDA blog post | Google
  98. [98] Training modality. LaMDA announcement describing dialog-only model
  99. [99] Active parameters. Meta Llama 3.1 model card
  100. [100] Context window. meta-llama/Llama-3.1 model card
  101. [101] Vision. meta-llama/Llama-3.1 model card
  102. [102] Serving mode. meta-llama/Llama-3.1 model card
  103. [103] Training parameters. Meta Llama 3.1 model card
  104. [104] Training modality. meta-llama/Llama-3.1 model card
  105. [105] Context window. meta-llama/Llama-3.1 model card
  106. [106] Vision. meta-llama/Llama-3.1 model card
  107. [107] Serving mode. meta-llama/Llama-3.1 model card
  108. [108] Training modality. meta-llama/Llama-3.1 model card
  109. [109] Context window. meta-llama/Llama-3.1 model card
  110. [110] Vision. meta-llama/Llama-3.1 model card
  111. [111] Serving mode. meta-llama/Llama-3.1 model card
  112. [112] Training modality. meta-llama/Llama-3.1 model card
  113. [113] Active parameters. OPT release
  114. [114] Context window. Meta AI blog
  115. [115] Vision. Meta AI blog
  116. [116] Serving mode. Meta AI blog
  117. [117] Training parameters. OPT release
  118. [118] Training tokens. Meta OPT release blog
  119. [119] Training regime. OPT release blog | Meta
  120. [120] Training modality. OPT release blog describing dense decoder transformer
  121. [121] Active parameters. NVidia & Microsoft announcement
  122. [122] Context window. Microsoft product page
  123. [123] Vision. Microsoft announcement
  124. [124] Serving mode. Microsoft product page
  125. [125] Training parameters. NVidia & Microsoft announcement
  126. [126] Training tokens. Microsoft Azure AI announcement
  127. [127] Training regime. Megatron-Turing NLG release
  128. [128] Training modality. Azure AI service introduction describes a text-only transformer
  129. [129] Active parameters. Mistral model overview
  130. [130] Context window. Codestral - Mistral Docs
  131. [131] Vision. Codestral - Mistral Docs
  132. [132] Serving mode. Codestral - Mistral Docs
  133. [133] Training parameters. Mistral model overview
  134. [134] Training modality. Codestral - Mistral Docs
  135. [135] Context window. Research-licensed edge model - Mistral Docs
  136. [136] Vision. Research-licensed edge model - Mistral Docs
  137. [137] Serving mode. Research-licensed edge model - Mistral Docs
  138. [138] Training modality. Research-licensed edge model - Mistral Docs
  139. [139] Context window. Research-licensed edge model - Mistral Docs
  140. [140] Vision. Research-licensed edge model - Mistral Docs
  141. [141] Serving mode. Research-licensed edge model - Mistral Docs
  142. [142] Training modality. Research-licensed edge model - Mistral Docs
  143. [143] Active parameters. Mistral Large 2 launch note
  144. [144] Context window. Mistral Large 2.1 - Mistral Docs
  145. [145] Vision. Mistral Large 2.1 - Mistral Docs
  146. [146] Serving mode. Mistral Large 2.1 - Mistral Docs
  147. [147] Training parameters. Mistral Large 2 launch note
  148. [148] Training modality. Mistral Large 2.1 - Mistral Docs
  149. [149] Context window. Mistral Small 3.1 - Mistral Docs
  150. [150] Vision. Mistral Small 3.1 - Mistral Docs
  151. [151] Serving mode. Mistral Small 3.1 - Mistral Docs
  152. [152] Training modality. Mistral Small 3.1 - Mistral Docs
  153. [153] Active parameters. Nexos ChatGPT version history
  154. [154] Context window. OpenAI ChatGPT blog
  155. [155] Vision. OpenAI ChatGPT blog
  156. [156] Serving mode. OpenAI ChatGPT blog
  157. [157] Training parameters. Nexos ChatGPT version history
  158. [158] Training regime. OpenAI ChatGPT blog
  159. [159] Training modality. OpenAI ChatGPT blog
  160. [160] Training hardware. OpenAI ChatGPT blog
  161. [161] Active parameters. Exploding Topics GPT-4 parameters report
  162. [162] Context window. OpenAI GPT-4 research release
  163. [163] Vision. OpenAI GPT-4 research release
  164. [164] Serving mode. OpenAI GPT-4 research release
  165. [165] Training parameters. Exploding Topics GPT-4 parameters report
  166. [166] Training tokens. Assumed same proxy as GPT-5 series
  167. [167] Training regime. OpenAI GPT-4 technical report
  168. [168] Training modality. OpenAI GPT-4 research release
  169. [169] Training hardware. OpenAI GPT-4 research release
  170. [170] Active parameters. Internal approximate profile for extrapolation
  171. [171] Context window. GPT-4o mini model docs | OpenAI API
  172. [172] Vision. Introducing GPT-4o mini | OpenAI
  173. [173] Serving mode. GPT-4o mini model docs | OpenAI API
  174. [174] Training parameters. Internal approximate profile for extrapolation
  175. [175] Training regime. Introducing GPT-4o mini | OpenAI
  176. [176] Training modality. Introducing GPT-4o mini | OpenAI
  177. [177] Training hardware. Introducing GPT-4o mini | OpenAI
  178. [178] Active parameters. Artificial Analysis size class for GPT-5 mini (medium = 40B–150B); midpoint used as screening estimate
  179. [179] Context window. GPT-5 mini Model | OpenAI API
  180. [180] Vision. GPT-5 mini Model | OpenAI API
  181. [181] Serving mode. GPT-5 mini Model | OpenAI API
  182. [182] Training parameters. Artificial Analysis size class for GPT-5 mini (medium = 40B–150B); midpoint used as screening estimate
  183. [183] Training modality. GPT-5 mini Model | OpenAI API
  184. [184] Active parameters. Artificial Analysis size class for GPT-5 nano (medium = 40B–150B); midpoint used as screening estimate
  185. [185] Context window. GPT-5 nano Model | OpenAI API
  186. [186] Vision. GPT-5 nano Model | OpenAI API
  187. [187] Serving mode. GPT-5 nano Model | OpenAI API
  188. [188] Training parameters. Artificial Analysis size class for GPT-5 nano (medium = 40B–150B); midpoint used as screening estimate
  189. [189] Training modality. GPT-5 nano Model | OpenAI API
  190. [190] Active parameters. Alan D. Thompson GPT-5 estimate (~300B)
  191. [191] Context window. GPT-5.2 Model | OpenAI API
  192. [192] Vision. GPT-5.2 Model | OpenAI API
  193. [193] Serving mode. GPT-5.2 Model | OpenAI API
  194. [194] Training parameters. Alan D. Thompson GPT-5 estimate (~300B)
  195. [195] Training modality. GPT-5.2 Model | OpenAI API
  196. [196] Context window. GPT-5.2 pro Model | OpenAI API
  197. [197] Vision. GPT-5.2 pro Model | OpenAI API
  198. [198] Serving mode. GPT-5.2 pro Model | OpenAI API
  199. [199] Training modality. GPT-5.2 pro Model | OpenAI API
  200. [200] Active parameters. gpt-oss-120b Model | OpenAI API
  201. [201] Context window. gpt-oss-120b Model | OpenAI API
  202. [202] Vision. gpt-oss-120b Model | OpenAI API
  203. [203] Serving mode. gpt-oss-120b Model | OpenAI API
  204. [204] Training parameters. gpt-oss-120b Model | OpenAI API
  205. [205] Training modality. gpt-oss-120b Model | OpenAI API
  206. [206] Active parameters. gpt-oss-20b Model | OpenAI API
  207. [207] Context window. gpt-oss-20b Model | OpenAI API
  208. [208] Vision. gpt-oss-20b Model | OpenAI API
  209. [209] Serving mode. gpt-oss-20b Model | OpenAI API
  210. [210] Training parameters. gpt-oss-20b Model | OpenAI API
  211. [211] Training modality. gpt-oss-20b Model | OpenAI API
  212. [212] Active parameters. Open Release of Grok-1 | xAI
  213. [213] Context window. https://docs.x.ai/developers/models
  214. [214] Vision. https://docs.x.ai/developers/models
  215. [215] Serving mode. https://docs.x.ai/developers/models
  216. [216] Training parameters. Open Release of Grok-1 | xAI
  217. [217] Training regime. Open Release of Grok-1 | xAI
  218. [218] Training modality. https://docs.x.ai/developers/models
  219. [219] Active parameters. Grok-2 parameter-count estimate from open-weight config discussion
  220. [220] Training parameters. Grok-2 parameter-count estimate from open-weight config discussion
  221. [221] Training regime. Grok 2 update | xAI
  222. [222] Active parameters. Alan D. Thompson Models Table estimate for Grok 4 (600B)
  223. [223] Context window. Models and Pricing | xAI
  224. [224] Vision. Models and Pricing | xAI
  225. [225] Serving mode. Models and Pricing | xAI
  226. [226] Training parameters. Alan D. Thompson Models Table estimate for Grok 4 (600B)
  227. [227] Training modality. Models and Pricing | xAI

Country factors for carbon and water recalculation

No. Country Year Carbon intensity Reference
1 France 2024 40 gCO2e/kWh ImpactLLM method note: project screening electricity factors
Project screening default retained for rapid comparative estimation; not an audited national inventory factor.
2 United States 2024 385 gCO2e/kWh ImpactLLM method note: project screening electricity factors
Project screening default retained for rapid comparative estimation; not an audited national inventory factor.
3 Germany 2024 380 gCO2e/kWh ImpactLLM method note: project screening electricity factors
Project screening default retained for rapid comparative estimation; not an audited national inventory factor.
4 United Kingdom 2024 180 gCO2e/kWh ImpactLLM method note: project screening electricity factors
Project screening default retained for rapid comparative estimation; not an audited national inventory factor.
5 Canada 2024 120 gCO2e/kWh ImpactLLM method note: project screening electricity factors
Project screening default retained for rapid comparative estimation; not an audited national inventory factor.
6 China 2024 540 gCO2e/kWh ImpactLLM method note: project screening electricity factors
Project screening default retained for rapid comparative estimation; not an audited national inventory factor.