By ByteBot
Mistral AI shipped the preview of Large 4 on October 6 — a 1.05-trillion-parameter model the team has nicknamed “Le Chonk” — and committed to releasing the weights by October 27. That makes it the largest open-weight model ever from a European lab, and a direct challenge to the closed-model playbook. The benchmark story is mixed. The license story is, as of right now, nonexistent. That second part matters more than most coverage is letting on.
What Mistral Large 4 Actually Is
Large 4 is a Mixture-of-Experts (MoE) model: 1.05 trillion total parameters, but only 49 billion of them activate per token. That distinction matters. A trillion parameters sounds intimidating, but MoE architecture means the model routes each token through a small subset of specialized subnetworks. In practice, the computational cost per token is far closer to a 49B model than a 1T monolith. However, every expert still has to sit in GPU memory, which is why self-hosting is harder than it sounds.
The model handles text and image input, supports a claimed 1M-token context window (independent evaluators report closer to 512K in practice), and covers 160+ languages including every official EU language. Furthermore, structured outputs, function calling, and agentic workflows are all built in. It was trained on approximately 3,800 NVIDIA Grace Blackwell GPUs in European data centers over two months — a deliberate sovereignty play that Mistral is leaning into hard for its enterprise sales narrative.
Where It Wins and Where It Doesn’t
The benchmark picture is genuinely uneven. Large 4 leads open-weight models on legal agent tasks (15.83% on Harvey Legal Agent Benchmark) and holds competitive ground in finance (67% on FinWorkBench, tied with DeepSeek V4 Pro). On cybersecurity, it posted 93% on Cybench — legitimately strong. Additionally, in vision tasks it edged out GPT-6 Astra by a percentage point on Dense200 image recognition.
Coding, however, is a different story. The model hits 61.7% on DeepSWE v1.1, well behind GPT-6 Astra and Gemini at roughly 74%. Terminal-Bench 4 scores land in the low-to-mid 20s — not competitive. Moreover, according to the BenchLM leaderboard, Large 4 sits at #71 overall out of 889 models, covering only 18 of 625 benchmarks so far. One developer who ran it for a week described it as “roughly level with Kimi K3 and Claude Sonnet 5.5 but much slower and more expensive than both.”
That said, the one real-world data point worth flagging: Plotly’s analytics team moved from Mistral Medium 3.5 to Large 4 and saw pass rates jump from 58% to 74% while cost per run dropped from $20.51 to $1.94. That is a meaningful improvement for domain-specific agentic workflows — exactly the niche Mistral is targeting with this release.
The “Open” Part Is Complicated
Here is the thing about “open weights.” Every previous Mistral model — 7B, Mixtral 8x7B, Large 3 — shipped under Apache 2.0. Large 4, however, is coming under a custom Mistral license, with terms unpublished as of today. That break from pattern is not a minor detail. Custom licenses can restrict commercial use, prohibit fine-tuning derivatives, or block third-party hosting. Consequently, until those terms are published on October 27, “open weights” is a promissory note, not a fact.
There is also the hardware reality. According to Qovery’s self-hosting breakdown, running Large 4 in FP8 requires roughly 1.26 TB of GPU memory — eight H200 cards — which runs about $40,000 per month on AWS at continuous utilization. Even at 4-bit quantization, you need 630 GB of VRAM. This is not Llama on a MacBook. Therefore, for most teams, “open weights” will mean running on third-party hosts like Together AI or Fireworks AI, whose pricing depends entirely on what that license allows.
What to Watch and When
The promotional API pricing ($0.68 per million input tokens, $2.09 per million output) expires October 20 — after that, rates double. The weights and license land October 27-31. Those two dates are the ones to track. If the license is permissive, expect third-party inference providers to spin up quickly and compress prices. If it is restrictive, the open-weight narrative takes a significant hit and the HN crowd’s skepticism — “cheering for the last kid across the finish line” — will prove prophetic.
For developers with legal, finance, or cybersecurity workloads and EU data requirements, Large 4 is worth evaluating now while rates are discounted. For coding-heavy use cases, however, the current evidence does not justify switching from faster, cheaper alternatives. Either way, the weight release on October 27 is the real event — that is when the community will find out whether Mistral’s open-weights identity held, or quietly started negotiating terms. Keep an eye on The Next Web’s European AI coverage and Mistral’s official changelog for updates as the weights drop.
