ModelsArchitectures & capability
Mistral previews Large 4, a 1-trillion-parameter model, with open weights planned for 27 October
Mistral has opened an API preview of Large 4, a mixture-of-experts model with about 1 trillion parameters, 49 billion of them active per token. An independent rating shows a big jump over its predecessor, but it trails leading rivals on coding.
AT Pulse research deskOctober 7, 2026 · 4 min read · AI-assisted analysisMistral Large 4 is a sparse mixture-of-experts model with about 1 trillion total parameters and 49 billion active per token. It takes images and text as input and produces text. The API preview opened on 6 October 2026, priced at $1.36 per million input tokens and $4.18 per million output tokens. Weights are planned for 27 October under a custom Mistral license. [5] [3] [6]
The main independent signal so far is Artificial Analysis, which gives Large 4 a score of 38. Mistral Large 3 scored 9, and Large 4 sits just behind DeepSeek 4.1 Flash. Simon Willison estimates it is roughly six months behind the frontier. [7]
Large 4 does not lead on coding. Mistral reports about 62% on DeepSWE. GLM-5.3 and Kimi K3 score around 69%, and leading closed models score around 74%. Most other benchmarks come from Mistral itself, and test setups differ across sources, so the scores are hard to compare. [5] [6]
Mistral is holding back the weights for about three weeks of safety testing because the model scores highly on offensive-cyber tasks. During that time it is red-teaming with cybersecurity firms and government authorities. Its strong cyber results and attack-resistance figures are self-reported. [1] [5]
one independent rating shows a large gain over Large 3, while Large 4's coding scores trail rivals.
Read the full assessment
Implication: buyers who want European-trained models with auditable weights may get a 1T-scale option, but decisions should wait for the weights and independent re-testing.
Executive brief
Mistral released a public API preview of Mistral Large 4 ("Le Chonk") on 6 October 2026. It is a mixture-of-experts model with about 1 trillion total parameters, 49B of them active per token. The weights are scheduled for release on 27 October (TNW). The most notable number so far comes from outside Mistral: Artificial Analysis gives it 38, compared with 9 for Mistral Large 3 (Simon Willison). That is a big jump for a European lab. Still, Mistral's headline DeepSWE score of about 62% is below the 69–74% posted on the public leaderboard (VentureBeat).
What changed and event timeline
Mistral Large 3 ships
Willison notes that this predecessor scored 9 on Artificial Analysis and calls its pelican-on-a-bicycle drawing "terrible." That low score is the baseline for ML4's gain.
€3B Series D led by Samsung
The round valued Mistral at more than €21B post-money, the largest equity raise by a European tech company. The CEO said the money will go to Mistral's own data centres and compute.
ML4 public preview
The API model ID is
mistral-large-4-0. Mistral calls it a natively multimodal, hybrid instruct-and-reasoning MoE and promises weights "end of this month."First external score
Artificial Analysis rates it 38, just behind DeepSeek 4.1 Flash. Willison estimates it is about six months behind the frontier.
(planned): Weights release
The release follows about three weeks of safety testing. Mistral says it will work with "trusted partners and governments" in the meantime.
More detail
Capabilities and access
- Model: Mistral Large 4, API ID
mistral-large-4-0. Image and text input, text output (VentureBeat). - Price: $1.36 per million input tokens and $4.18 per million output tokens, through Mistral Studio (Mistral).
- Reasoning: The API offers only two levels, "none" and "high" (Simon Willison).
Read the full section
- Model: Mistral Large 4, API ID
mistral-large-4-0. Image and text input, text output (VentureBeat). - Price: $1.36 per million input tokens and $4.18 per million output tokens, through Mistral Studio (Mistral).
- Reasoning: The API offers only two levels, "none" and "high" (Simon Willison).
- Languages: More than 160, including all official EU languages (Mistral).
- Weights: Planned under a custom Mistral license (VentureBeat).
- Target uses: Cybersecurity, finance and chip design (TechCrunch).
Technical analysis for researchers and developers
Read the full section
- Architecture: A sparse MoE with roughly 1T total and 49B active parameters (Mistral). Some outlets give the total as 1.05T (TNW). The expert count, routing and context length are not documented in the reviewed sources.
- Training compute: Mistral's post says 3,800 Grace Blackwell GPUs. TechCrunch and TNW report about 4,000, used for about two months at roughly 10 MW (TNW).
- Post-training: Reinforcement learning at about 33B tokens per day (Mistral).
- Reproducibility: Nobody can reproduce the results until the weights ship. Benchmark configurations also differ across sources, which makes scores hard to compare (VentureBeat).
Claims and evidence
- DeepSWE v1.1 61.7%, Terminal Bench 4.0 28.3%, Cybench 93% () — Reported by Mistral
- FinWorkBench 67%, DIOR-RSVG 73% vs GPT-6 Astra 68% ()
- Human evaluations by Surge AI and vals.ai ()
Read the full section
| Claim | Status |
| DeepSWE v1.1 61.7%, Terminal Bench 4.0 28.3%, Cybench 93% (Mistral) | Reported by Mistral |
| FinWorkBench 67%, DIOR-RSVG 73% vs GPT-6 Astra 68% (TNW) | Reported by Mistral; Mistral calls them preliminary |
| Human evaluations by Surge AI and vals.ai (Mistral) | Third-party, but commissioned by Mistral and described only by Mistral |
| Artificial Analysis score of 38 (Simon Willison) | Independent; Artificial Analysis's own model page was not among the reviewed sources |
| "Best in class" among open-weight models (TechCrunch) | An aspiration; Mistral says benchmark results are still pending |
Context and prior work
- ML4 succeeds Mistral Large 3 (December 2025) (Simon Willison).
- Mistral's smaller March 2026 model, Mistral Small 4 (Reasoning), scored 20 on Artificial Analysis (Artificial Analysis).
- On GPU use, Mistral says it trained with two to three times fewer GPUs than Chinese rivals.
Read the full section
- ML4 succeeds Mistral Large 3 (December 2025) (Simon Willison).
- Mistral's smaller March 2026 model, Mistral Small 4 (Reasoning), scored 20 on Artificial Analysis (Artificial Analysis).
- On GPU use, Mistral says it trained with two to three times fewer GPUs than Chinese rivals. That is a company claim; no comparison data was published (TechCrunch).
- Mistral's backers include ASML, which led the Series C, and Samsung, which led the Series D (Euronews).
Limitations, safety and contested findings
- Coding: On DeepSWE, GLM-5.3 and Kimi K3 score about 69%. So ML4 does not lead on coding (VentureBeat).
- Competitor scores: Many of the rival scores Mistral cites have not been independently checked (VentureBeat).
- Safety: Red-teaming with cybersecurity firms and state authorities is still under way.
Read the full section
- Coding: On DeepSWE, GLM-5.3 and Kimi K3 score about 69%. GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 score about 74%. So ML4 does not lead on coding (VentureBeat).
- Competitor scores: Many of the rival scores Mistral cites have not been independently checked (VentureBeat).
- Safety: Red-teaming with cybersecurity firms and state authorities is still under way. Mistral reports a 93.3% refusal rate on the B3 attack-resistance test, which is its own figure (Mistral).
- Misuse risk: The model has strong offensive-cyber scores, which is why Mistral is staging the weights release (TechCrunch).
Business and practitioner implications
- Sovereignty: Buyers who want models trained in Europe, with weights they can audit, get a credible option at the 1T scale (TechCrunch).
- Cost: At $1.36/$4.18 per million tokens, it is worth testing for finance, cyber and visual-grounding work (Mistral).
- Coding agents: Leading closed and Chinese models still score higher on DeepSWE (VentureBeat).
Read the full section
- Sovereignty: Buyers who want models trained in Europe, with weights they can audit, get a credible option at the 1T scale (TechCrunch).
- Cost: At $1.36/$4.18 per million tokens, it is worth testing for finance, cyber and visual-grounding work (Mistral).
- Coding agents: Leading closed and Chinese models still score higher on DeepSWE (VentureBeat).
- Self-hosting: Running a model with 1T total parameters takes multi-GPU servers, even though only 49B parameters are active per token.
- Procurement: Treat the 27 October weights release and independent re-evaluation as the point to decide.
Sources
Read the full section
- TechCrunch: Mistral's new 1T model aims to leapfrog closed and open rivals
- Mistral: Mistral Large 4
- VentureBeat: Mistral debuts Large 4 "Le Chonk"
- TNW: Mistral releases Large 4
- Simon Willison: Introducing Mistral Large 4: Le chonk
- Euronews: Mistral AI raises record €3 billion
- Artificial Analysis: Mistral Small 4
The source trail.
Sources (10)
Mistral’s new 1T model aims to leapfrog closed and open rivals
Article text retrieved; extracted text may omit tables or interactive elements.
techcrunch.comUnite.AI – Mistral Le Chonk
Related coverage; assess separately
unite.aiTNW: Mistral releases Large 4
thenextweb.comCNBC – Mistral
Related coverage; assess separately
cnbc.comMistral: Mistral Large 4
mistral.aiVentureBeat: Mistral debuts Large 4 "Le Chonk"
venturebeat.comSimon Willison: Introducing Mistral Large 4: Le chonk
simonwillison.netMistral Large 4 "Le Chonk": a 1.05T-parameter open-weight multimodal MoE
Related coverage; assess separately
marktechpost.com