1.8× less energy than Codestral.
Codestral Environmental Impact
Mistral's specialised code generation model
One query, in context
Energy, carbon and water for a single use. Code completion (~500 tokens input/output)
About 2.7× a Google search and 96% of a minute of laptop use.
33% of a Google search.
About 16× a drop of water and 16% of a teaspoon of water.
- Where it runsMistral AI (France)
- Grid carbon intensity50 g CO₂/kWh · 90% renewable
- Compared with a Google search3x more than a search (0.3 Wh)
- Water~1,250 queries to fill 1 litre
Where Codestral sits
#135 of 199 by energy per query. Heavier than 67% of all models. Lit dots are the same kind of model; click any dot to open it.
Want something lighter?
4.0× less energy than Codestral.
What a month of use adds up to
Pick how much you use it. Based on one quick use at a time, five days a week.
- Energy
- 173.2 Wh
- Carbon
- 8.7 g
- Water
- 173 mL
A month of 10 requests a day, 5 days a week. The energy is about 1.6× a kettle boil and 2% of a UK home's day.
Work out your own →- Energy
- 692.8 Wh
- Carbon
- 34.6 g
- Water
- 693 mL
A month of 40 requests a day, 5 days a week. The energy is about 6.3× a kettle boil and 9% of a UK home's day.
Work out your own →- Energy
- 3.46 kWh
- Carbon
- 173.2 g
- Water
- 3.46 L
A month of 200 requests a day, 5 days a week. The energy is about 31× a kettle boil and 47% of a UK home's day.
Work out your own →How does Codestral compare?
Next to similar models. Bars use a log scale, so each step is a big jump.
Detailed breakdown
Energy
Codestral is Mistral's dedicated code model, trained on 80+ programming languages. At ~0.8 Wh per code completion, it sits between lightweight autocomplete models and full-capability chat models. Code generation tends to produce longer outputs than chat, increasing per-query energy.
Power source & carbon
Hosted on Mistral's infrastructure in France, benefiting from one of Europe's cleanest grids at approximately 50 g CO2/kWh due to France's nuclear energy base.
Water use
At ~0.8 mL per completion on Mistral's French-grid reference (~1.0 L/kWh), Codestral's water footprint is low for a medium-sized model.
About Codestral
Codestral is a code assistant model from Mistral AI, released in May 29, 2024. Mistral's specialised code generation model. Each query uses 0.80 Wh of energy and produces 0.04 g of CO₂. That's about 2.7x what a Google search consumes.
Codestral benefits from running in France, one of the cleaner grid regions in our dataset at 50 gCO₂/kWh with 90% renewable energy. The same model running in a coal-heavy region would produce significantly more carbon per query.
These figures are estimates derived from hardware specifications and API benchmarks — Mistral AI has not published official energy data for Codestral. Actual consumption may vary significantly depending on batching, quantisation, and infrastructure optimisations that we cannot observe from outside.
Codestral in context
Key insights
- Runs on a 90% renewable grid — among the cleanest AI inference locations
What does your Codestral usage cost the planet?
Use the calculator to estimate your personal footprint based on how often you use Codestral.
Frequently asked questions
How much energy does Codestral use per query?
Each Codestral query consumes approximately 0.80 Wh of energy. This is 3x more than a traditional Google search (~0.3 Wh).
What is Codestral's carbon footprint?
Based on the carbon intensity of Mistral AI (France), each query produces approximately 0.04 g of CO2. The grid in this region has a carbon intensity of 50 g CO2/kWh with 90% renewable energy.
How much water does Codestral use?
Each query consumes approximately 0.80 mL of water, primarily used for cooling the data centers that process the request.
How does Codestral compare to a Google search?
A Codestral query uses 3x more than a Google search in terms of energy. A Google search uses approximately 0.3 Wh, while Codestral uses 0.80 Wh.
Technical details
- Architecture
- Dense Transformer (code-specialised)
- Parameters
- 22B
- Context window
- 32,000 tokens
- Release date
- 2024-05-29
- Open source
- No
- Training data cutoff
- 2024-05