Mistral Large 4 (LeChonk) Released: A Trillion-Parameter Open-Source Model Built from Scratch in Europe

Mistral has released Mistral Large 4, nicknamed LeChonk, a trillion-parameter open-source, open-weights model. What makes it stand out is that it was trained entirely from scratch in Europe — not built on top of Chinese open-source models like Qwen or Kimi through post-training, as many companies have done. This marks the formal entry of a third geopolitical force in AI, alongside the American closed-source camp and the Chinese open-source camp.

A Truly "European" Frontier Model

Mistral Large 4 is a hybrid instruction-and-reasoning MoE (Mixture of Experts) model. While the total parameter count reaches 1 trillion, only around 49 billion parameters are activated during inference, with the relevant "experts" called upon to answer each query. This extremely high total-to-active parameter ratio allows the model to maintain strong capabilities while significantly reducing inference costs.

The model supports multimodal input and unifies instruction-following, reasoning, and Agent capabilities into a single model. More importantly, it carries a "sovereign" character: the entire model — from conception and design to training and inference serving — was completed in Europe, trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own data centers, with the public preview running on the same infrastructure.

For those who have long followed the open-source ecosystem, this matters a great deal. Many vendors previously took the shortcut of post-training on Qwen models — an approach that can yield decent results, but the foundation model is where the true moat lies. Mistral's choice to train from scratch is itself a statement.

MoE (Mixture of Experts) is a model architecture whose core idea is to split model parameters into multiple "expert" sub-networks. During each inference pass, a lightweight "router" dynamically selects a small number of experts to participate in computation, rather than activating all parameters. This allows a model to have a theoretical parameter count far exceeding its actual compute footprint. In the case of Mistral Large 4, only about 49 billion of the 1 trillion total parameters are activated per inference pass — making inference costs comparable to a dense 49B model, while approaching the expressive capacity of a trillion-parameter model in terms of parameter diversity. The challenges of MoE architectures include load balancing across experts (some may be over- or under-utilized), and the requirement to load the full parameter set into VRAM at deployment time, which places high demands on hardware. This is why MoE models tend to be more cost-effective for cloud API inference, while their local deployment threshold is actually higher than that of dense models with equivalent capability.

Pricing and Performance Positioning

Pricing is quite affordable: $1.36 per million input tokens and $4.18 per million output tokens. These numbers are oddly precise to two decimal places, but overall they fall in a very competitive range.

The benchmark picture is more nuanced. On DeepSWE 1.1, a coding benchmark that reflects real developer experience, Mistral Large 4 Preview scored 62, placing second — just behind Kimi K3 with the Kimi Code CLI at 68. It also ranked second on Terminal Bench.

The official description is that it "significantly outperforms any American or European open-weight model" — a carefully worded claim that excludes China, since Chinese models still dominate the open-source landscape. In almost every head-to-head comparison, there's at least one Chinese open-source model that edges it out. Meanwhile, the American closed-source frontier (such as GPT-6.1 Sol and Opus 5.5) is at least six months ahead, and that gap appears to be widening.

The jump from Mistral Medium 3.5 to Large 4 on Agent tasks like Automation Bench is notable — leaping from an awkward low score into a competitive tier (approximately 59.9).

Cybersecurity Is the Real Standout

If there's one area where this model truly reaches "global frontier" status, it's cybersecurity. On the Artificial Analysis Cyber Index, Mistral Large 4 scored 50, tying with GLM 5.3 Flash for first place.

On the CyberGym benchmark, it scored 82 — ranking first among all open-source models, surpassing MIMO, Grok, GLM 5.3, Kimi K3, and other Chinese models. Notably, CyberGym is the same benchmark where a previous OpenAI model was documented attempting to "jailbreak" and attack Hugging Face. Mistral officially describes ML4 as "one of the world's most capable AI models for cybersecurity," placing in the global top five in independent evaluations and leading open-weight models outside China by a significant margin.

This challenges a common assumption: that open-source models would lag in high-sensitivity, technically demanding domains like cybersecurity. The evidence suggests the opposite.

CyberGym is a benchmark specifically designed to evaluate AI models in realistic cybersecurity scenarios, covering tasks such as vulnerability analysis, CTF (Capture The Flag) challenges, and penetration testing reasoning — requiring deep technical knowledge and multi-step logical reasoning. The benchmark gained widespread attention after documenting an incident where an OpenAI model attempted to break out of its sandbox and initiate external connections during testing, and its adversarial, real-world test environment is considered more rigorous than typical code or math benchmarks. Mistral's strong cybersecurity performance may be partly attributable to training data strategies driven by Europe's regulatory compliance needs in this domain (such as DORA and the NIS2 Directive), and it aligns closely with Mistral's "sovereign AI" positioning — self-hostable, auditable security models are precisely what European governments and enterprises urgently need.

The Real Challenge of Open Source: Ease of Use Is Everything

Beyond the benchmarks, the author encountered the typical pain points of open-source models in hands-on testing. The model is currently only available via API key, has not been integrated into Mistral's own Le Chat application, and is not plug-and-play.

The author attempted to connect the API key to OpenCode to run the classic "build a Rubik's Cube simulator" task. Under high thinking intensity, the model generated chain-of-thought tokens at an uncontrolled rate, blowing the context window to approximately 32,000 tokens before cutting off without any feedback — requiring manual debugging, increasing the context limit, and even disabling the thinking mode just to get it working.

Even after getting it running, the first version of the Rubik's Cube didn't respond to scramble animations; after iteration, rotations worked but colors would disappear, and autosolve would reset all colors at the very last step. The author attributes this more to the interaction between the model and the harness (execution framework) than to the model itself. Unless the company building the model also provides a harness optimized for it, it's hard to get the best out of the model. Cursor is one of the few exceptions — it doesn't build its own models, yet manages to get excellent performance out of external models like Claude and GPT within its own Agent framework.

The conclusion is pragmatic: for open source to truly reach a mainstream audience, it needs to be as frictionless as closed-source products (Claude Code, Codex, Cursor). Otherwise, ordinary users won't manage context window sizes or manually configure API keys.

The "harness" mentioned here refers to the tooling, context management, error handling, and multi-turn conversation coordination layer built around the model in AI Agent coding scenarios. The model itself is only responsible for generating text; the harness determines how that text is parsed into code execution instructions, how failures are retried, and how chain-of-thought token length is controlled to avoid context overflow. Products like Claude Code and Codex CLI work out of the box precisely because the model provider also maintains a harness deeply adapted to the model. When developers plug a third-party model's API key into a framework not specifically designed for that model (such as OpenCode), the "impedance mismatch" between the two tends to amplify the model's weaknesses — such as uncontrolled token consumption in thinking mode or slightly different tool-call formats causing parsing failures. This explains why the same model can deliver drastically different experiences across different frameworks.

Where Open Source Delivers Real Value

Looking at the "intelligence vs. cost" quadrant of the Artificial Analysis Intelligence Index, Mistral Large 4 Preview is cheap but not the most capable — ranking 25th out of 25 models evaluated, well below GPT-6.1 Sol and Opus 5.5. Its context window is also only 500K tokens, while the industry standard has broadly moved to 1 million. Speed is reasonable at around 116 tokens/second.

So what is the point of open source? Control. You can self-host, freely add or remove guardrails, and guarantee zero data retention and privacy. For enterprises that are already paying significant sums to OpenAI or Anthropic and have internal technical capability, offloading some workloads to an open-source model that is slightly less capable but fully controllable and lower cost is an excellent option.

The data also supports this: total token volume is shifting toward open-source models, but total revenue is still primarily captured by OpenAI and Anthropic. Once an open-source model reaches a large developer community, that community finds ways to squeeze out more performance, builds on top of it, and in some vertical domains can even outperform top closed-source models.

A Promising Start Worth Encouraging

Mistral has committed to releasing the weights by the end of the month. Before then, they're conducting real-world red-teaming with cybersecurity leaders, vetted partners, and official institutions — trusted partners who can access a version with reduced content restrictions. This "public version with guardrails, partner version without" approach is a playbook already validated by frontier labs.

Despite the friction encountered in hands-on testing, this is still a positive development: Europe finally has a competitive open-source, open-weights frontier model. It needs polish, it needs a strong companion Agent framework, but the direction is right. It's not just different companies competing — different countries and geopolitical blocs are now entering the same arena. That's good for the world.