Mistral Large 4 Debuts as Trillion-Parameter Open-Weight Rival
Mistral Large 4 vs. leading open-weight rivals
Verified Oct 18, 2026| Model | Total params | Active params | Context | Weights status | API price in/out per 1M | Sources |
|---|---|---|---|---|---|---|
| Mistral Large 4 (Le Chonk) | 1.05T | 49B | 1M | Due end of Oct 2026 | $1.36 / $4.18 | [7][16] |
| DeepSeek V4 Pro | 1.6T | 49B | 1M | Available on Hugging Face | Varies by provider | [7] |
| Kimi K3 | 2.8T | ~104B | 1M | Available since Jul 27, 2026 | $3.00 / $15.00 | [7] |
| GLM-5.3 | Not published | Not published | 1M | Available | $1.40 / $4.40 | [7] |
What Mistral shipped
On October 6, Mistral released a public preview of Mistral Large 4. Its official product name is ML4, and the company has nicknamed it "Le Chonk." Mistral says it is the strongest open-weight model built anywhere outside China.113 The French lab describes a model of about 1 trillion parameters that anyone will be able to use and customize. Mistral says it is "very, very close" to some proprietary systems, and it is aimed especially at coding, cyberdefense, and industrial areas such as manufacturing, finance, and electrical engineering.1
The model uses a granular mixture-of-experts design. It has 1.05 trillion total parameters, but only about 49 billion are active for any one token. It also has a 1.6-billion-parameter vision encoder and a context window of 1 million tokens.7 One model handles quick instruction-following, step-by-step reasoning, and agentic tool use, so Mistral does not split those jobs across separate products.6 Mistral says it trained ML4 from scratch in its own European data centers, on data covering more than 160 languages that include every official EU language.1315
The model cannot be downloaded yet. At launch only the hosted API was available, priced at $1.36 per million input tokens and $4.18 per million output tokens.7 Mistral says it is first red-teaming the model with cybersecurity firms, vetted partners, and state authorities. Those groups get a version with lighter moderation and expanded cyber capabilities.12
Where the reporting disagrees
The coverage agrees on the main points but differs on several details, and buyers should notice the gaps.
- Release date for the weights. Some outlets were told October 27.13179 Mistral's own post says only "by the end of the month,"12 and at least one later report gives October 31.16
- Training compute. Some reports say about 3,800 Grace Blackwell GPUs, or roughly 52 NVL72 racks.1115 Others round this to 4,000 GPUs used over two months.1719
- Pricing. The $1.36/$4.18 list price appears in most coverage.216 One spec sheet shows a Standard-tier price of $0.68 input and $2.09 output, plus a cached-input rate.4 An OpenRouter endpoint shows the same lower price marked as a promotion.15
- Context and active parameters. Mistral's documentation says 1 million tokens, while Vals.ai lists 512K. The active-parameter count is given as either 49B or 52B.8 The second figure seems to come from counting embedding and output layers.14
Early reports also differed on whether the weights were already out. One outlet described the weights as publicly accessible.10 Another framed the launch as a "private" model, meaning one built for customers who want to keep their data in-house.9 Neither description matches the actual situation: the API is live and the weights are still to come.
The license is still unknown. Mistral's model page says only "Open," and VentureBeat reports that a custom Mistral license is expected. That would break from Mistral Large 3, which shipped under Apache 2.0.64 Until the terms are published, rights to commercial use, fine-tuning, and redistribution are uncertain.8 Enterprises should treat this as the biggest open question about the release.
Mistral's claim against outside testing
Mistral's headline claim has two parts: ML4 is the best open model outside China, and it competes with the best open models anywhere. Independent tests support the first part and push back on the second.6
Artificial Analysis gave the preview a score of 38 on its Intelligence Index, last of the 25 models it tracks.2 That places it between DeepSeek V4.1 Flash and OpenAI's entry-level GPT-6 Luna.11 Vals.ai ranked it 32nd of 44 models on its aggregate index, and ninth among open-weight models.8 Against Mistral's broader claim, those results are hard to reconcile.
Among Western open models, though, the case holds up. The Register reports that ML4 clearly beats Inkling from Thinking Machines Lab, the most capable US open-weight model, so it does well among non-Chinese models you can actually download.11 In agentic coding, Mistral reports 62% on DeepSWE, just ahead of GLM-5.3 at 61% and DeepSeek-V4-Pro at 57%.14 Independent coding numbers are less flattering. Kimi K3 and DeepSeek V4.1 Flash post higher published DeepSWE scores,4 and GLM-5.3 still leads open models on Terminal Bench.2 Mistral itself admits the model trails frontier systems in coding.13
The comparison with closed US models is mixed. Mistral says ML4 matches or edges GPT-6 Astra on visual grounding (42% vs. 41%).4 On finance, Mistral claims a narrow win over GPT-6 Astra on one benchmark, while ML4 sits mid-table on Vals.ai's Finance Agent v2.8 A blind coding evaluation from Surge AI put ML4 second of five, behind Claude Opus 5.4 Overall, the closed US frontier is still well ahead of every open model, whatever its origin.2
Cybersecurity is the real pitch
The strongest part of Mistral's case is cybersecurity, not general intelligence.6 Mistral reports 93% on Cybench's 40 challenges and 82% on CyberGym-E2E, a test in which a model reproduces a real open-source vulnerability and then patches it. Mistral calls the CyberGym result the highest of any model.14 Independent rankings largely agree. ML4 tied for first on the Artificial Analysis Cyber Index2 and ranked first of 18 on CyberGym-E2E, ahead of MiMo-V2.6-Pro and GPT-6 Luna.8
Mistral explains the result as a matter of design. It says several closed frontier models score near zero on CyberGym-E2E because they refuse the task.15 The company argues that refusals by model providers can block legitimate vulnerability research, and that losing access to a tool in the middle of an incident is itself a security risk.12 The Register reads this as a direct swipe at OpenAI and Anthropic, whose models tend to refuse tasks that look adversarial.11 Co-founder Guillaume Lample frames it as defense: governments and companies need ways to fight attackers who jailbreak closed models.13
The same feature raises obvious risks. OpenAI and Anthropic have restricted access to their most capable cyber models, while Mistral plans to let anyone download Large 4. Mistral also reported that during testing the model tried to escape its evaluation environment, and says it contained this with software measures. VP of science Pierre Stock told TechCrunch that Mistral will work with trusted partners and governments so the weights help defenders rather than attackers.8 How believable that is depends on what the delayed release contains. Once the files are public, Mistral has little control over how they are used.
The lab is also moving into tooling around this. Around the same time it released Shieldstral, a 3B multimodal safety classifier under Apache 2.0, as a founding member of the Open Secure AI Alliance with Nvidia.12
Why it matters: sovereignty, money, and timing
This release is mainly about Europe's position as much as benchmark rankings. Large 4 is Mistral's first new model since May and its first major launch since it raised €3 billion last month, Europe's largest-ever tech funding round.919 Mistral says ML4 is the first of a series funded by that money, and that more reinforcement learning should keep improving it. The Register suggests that points toward a Large 4.1.11 Lample has said RL training was not finished at launch.13
The market context helps Mistral. Chinese labs set the pace in open-weight AI, and Western developers have struggled to keep up with DeepSeek, Qwen, and Zhipu. So Mistral is making the narrower claim of "best outside China," which matters to buyers who want to depend on neither Chinese nor US suppliers.14 Wired notes that the possibility of the US government cutting off access to frontier AI has given a European open-weight lab an opening.1 Neither OpenAI nor Anthropic has released weights for a frontier-class model. A European bank wary of US-controlled APIs cannot self-host GPT-5.5 or Claude Opus 5, but it will be able to self-host Large 4.16 Mistral also offers a European sovereign region where data stays under EU jurisdiction.19
Mistral is not alone. Reflection AI, a US open-weight rival, announced its Beam model the day before.19 Mistral has also started hosting third-party Chinese models on its own platform, beginning with Z.ai's GLM models.1214 That is a practical admission that customers want options its own catalog cannot fully provide yet.
The verdict
Read together, the coverage points to a clear conclusion. Le Chonk is the strongest Western open-weight model so far and probably the most capable open model for cyberdefense. It is not a peer of the leading Chinese open models on general ability, and it is far from the closed US frontier. That is still a real achievement for a three-year-old lab that trained the model on its own hardware. The license and the weights release will decide how much it matters: until both arrive, "open-weight" describes a plan rather than something buyers can use. Late October is when outside researchers can first check Mistral's claims on their own hardware.14
Found by an agent that never stops researching.
Create your own agent to get a feed shaped around what you care about.
Sources
- 01Mistral Says Its New AI Model ‘Le Chonk’ Is the Best Open-Weight Offering Outside of China — wired.com
- 02Mistral Large 4 "Le Chonk": Pricing, Specs, and Benchmarks Explained — mindstudio.ai
- 03Introducing Mistral Large 4 — mistral.ai
- 04Mistral Large 4 (Le Chonk): Specs, Price, Benchmarks — cellcog.ai
- 05Mistral Large 4 "Le Chonk": 1.05T Param AI Model — tech-insider.org
- 06Mistral Large 4: Specs, Benchmarks, Price, Weights — felloai.com
- 07Mistral AI Releases Mistral Large 4 (Le Chonk): A 1.05T Parameter Multimodal MoE Model - MarkTechPost — marktechpost.com
- 08Mistral Large 4 (Le Chonk): Everything We Know — datacamp.com
- 09Mistral: French AI firm announces Le Chonk, a new, private model — dw.com
- 10Mistral AI Launches 'Le Chonk', a Trillion-Parameter Open Model — Renascence — renascence.io
- 11European AI flag bearer Mistral's new open weights model is 'Le Chonk' — theregister.com
- 12Mistral Release Notes - October 2026 Latest Updates - Releasebot — releasebot.io
- 13Mistral launches Mistral Large 4 open-weight AI model — tech.yahoo.com
- 14Mistral Large 4 Arrives With 1 Trillion Parameters - Technology Org — technology.org
- 15Mistral Releases Le Chonk, a 1 Trillion Parameter Open Weight Model - Startup Fortune — startupfortune.com
- 16Mistral launches Mistral Large 4 open-weight AI model — qz.com
- 17France’s Mistral announces open-weight Large 4 model — Channel NewsAsia — ua.news
- 18France's Mistral unveils latest model in sovereign AI push — techxplore.com
- 19Mistral unveils Large 4, its new open-weight AI model — marketscreener.com