October 12, 2026 | AI | By Mahesh

Mistral AI opened a public preview of Mistral Large 4 on October 6, 2026. The model is a natively multimodal mixture-of-experts system with 1.05 trillion total parameters and 52 billion active parameters per token, and developers can call it today through Mistral Studio [1][2]. The open weights are promised by the end of October. Mistral nicknames the model "le Chonk" and sells it on coding, agent workflows and cybersecurity, where it argues that defenders need a model that does not refuse security work.

What is Mistral Large 4 and what is new

Mistral Large 4 (model id mistral-large-4, version v26.10) is the largest model Mistral AI has built. The docs describe it as an open-weight hybrid instruct and reasoning model that unifies instruction following, reasoning and agentic behavior in one set of weights, with fluency in more than 160 languages [2][1]. It sits at the top of the catalog above Mistral Large 3, which shipped under Apache 2.0.

The headline change is vision. A 1.6 billion parameter vision encoder lets the model read charts, engineering drawings and satellite imagery, and Mistral says it beats GPT-6 Astra on the Dense 200 visual grounding test, 42% to 41% [1]. Through Mistral Studio the model supports function calling, structured outputs, batch processing and the Agents API [2].

"by far the strongest open-weight model from the US or Europe"

Guillaume Lample, co-founder and chief scientist, Mistral AI, on X [10]

The release is the first milestone funded by Mistral's €3 billion Series D, which the company calls the largest equity round ever raised by a European technology company [1]. Depth Grid's Q3 2026 global venture funding analysis covers the wider flow of capital into AI labs this quarter.

Under the hood: parameters, training and hardware

Mistral Large 4 has 1.05 trillion total parameters and 52 billion active per token, which means about 5% of the weights fire on any given token [2]. Some outlets report 49 billion active parameters. Mistral's own announcement and documentation both say 52 billion, so this article uses that figure.

Mistral trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and the preview is served from the same infrastructure [1]. Post-training leans heavily on reinforcement learning. At roughly 3,000 GPUs a single run produces about 33 billion tokens a day, of which around 16 billion are trainable completion tokens after filtering [1].

Several architecture details are still missing. Mistral has not published the expert count, the routing scheme or the layer layout, and says these will arrive alongside the weights [11]. Hardware guidance is also absent. A simple estimate puts the weights alone at about 1.05 terabytes at 8-bit precision, before any memory for the key-value cache, so self-hosting means a multi-node GPU cluster. That figure is Depth Grid arithmetic, not a Mistral number.

Benchmarks: Mistral's claims against outside checks

On agentic coding, Mistral reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4. Its combined Coding Agent Index score of 49.8% places it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max [1]. In a blind human evaluation run with Surge AI, annotators scored the preview 3.74 on a scale of 1 to 5. That ranked second of five models, behind Claude Opus 5 at 4.22 and ahead of Kimi K3 at 3.59 and GLM-5.3 at 3.60 [1].

Cybersecurity is where Mistral makes its strongest claims. The model ranks in the top five on the Artificial Analysis Cyber Index and scores 82% on the index test that asks a model to reproduce a real open-source vulnerability and then patch it. It also solves 93% of Cybench, a set of 40 security competition exercises [1]. Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero on the reproduce-and-patch test because they refuse the task, so the 82% partly measures willingness to engage as well as skill [1].

On workplace agents the model scores 59.9% on AutomationBench, which covers 657 business workflows across Gmail, Google Sheets, Slack and Salesforce, and reaches 1,393 Elo on AA-Briefcase [1]. For safety, Mistral reports that it resists 93.3% of attacks on Lakera's B3 benchmark and scores 1.691 out of 2 on the KORA benchmark [5][6]. It also says the model refuses malicious cyber prompts more often than any other open-weight model it compared.

Outside views are less flattering. The Vals Index puts Large 4 at 48.05%, 32nd of 44 models and ninth among open-weight entries [8]. Mistral itself calls the benchmark results preliminary and expects them to change before the weights ship [9]. Most figures above were run or selected by the vendor, so treat them as claims until independent evaluators publish numbers on the final weights.

Pricing, availability and how to access it

Mistral Large 4 lists at $1.36 per million input tokens, $0.14 per million cached input tokens and $4.18 per million output tokens, with a 1 million token context window [2][3]. At the time of writing the docs page also shows a promotional rate of $0.68 input, $0.07 cached and $2.09 output, with no end date displayed. Developers can try the preview API on Mistral Studio [4].

The weights are due by the end of October, and The Next Web reports a date of October 27 [9]. As of today the official Mistral GitHub organization is the place to watch for the release [7]. The model page lists the license only as "Open", so the actual terms are unknown. Mistral also says it operates a European deployment end to end under European law, which matters for buyers with data residency rules [1].

Until the weights ship, Mistral is red-teaming the model with cybersecurity leaders, vetted partners and state authorities, who get the same model with reduced moderation and expanded cyber capabilities [1]. Cyber-capable models are arriving as Washington tightens reporting rules, which Depth Grid covered in White House orders AI firms to report incidents after Anthropic.

Limitations, trade-offs and alternatives

Mistral Large 4 cannot be self-hosted yet, which makes the open-weight label a promise rather than a fact today. Four gaps remain open: the license terms, the architecture details, independent benchmarks on the final weights and the hardware needed to run a 1.05 trillion parameter model. A preview model can also change behavior before general availability.

Among open-weight rivals, DeepSeek V4 Pro has weights on Hugging Face and Kimi K3 has been downloadable since July 27, 2026 [11]. GLM-5.3 from Z.ai is also in the comparison set Mistral uses. Closed models still lead on the human coding evaluation, with Claude Opus 5 ahead of Large 4 [1]. The open-weight field is also getting more crowded on the US side, and Depth Grid has tracked Nvidia's talks over Reflection AI, a lab The Next Web describes as having just unveiled its first open-weight model, Beam [9].

Sources

1. Mistral AI. "Introducing Mistral Large 4," October 6, 2026. mistral.ai
2. Mistral AI. "Mistral Large 4," model documentation, v26.10. docs.mistral.ai
3. Mistral AI. Pricing page. mistral.ai/pricing
4. Mistral AI. Mistral Studio, preview API access. console.mistral.ai
5. Lakera. "B3 Agent Security Benchmark," Hugging Face dataset. huggingface.co
6. KORA Benchmark. korabench.ai
7. Mistral AI. Official GitHub organization. github.com/mistralai
8. Technology.org. "Mistral Unveils Large 4, a 1-Trillion-Parameter Open-Weight Model," October 7, 2026. technology.org
9. The Next Web. "Europe's Mistral launches Large 4 to challenge China's lead in open AI models." thenextweb.com
10. BetaNews. "Mistral opens trillion-parameter Large 4 preview, weights due in October." betanews.com
11. MarkTechPost. "Mistral AI Releases Mistral Large 4 (Le Chonk)," October 6, 2026. marktechpost.com

Read More on Depth Grid

• Tech News Today, Oct 8: GPT-6, SpaceX, Nuclear Clocks
• SAP Buys TechWolf: What the AI Work Deal Means
• Volantis $88M Funding: Light-Based AI Chips Explained
• The AI Chip Market Everyone Forgot: Robots That Don't Talk

Article by Mahesh | Depth Grid