Mistral Large 4 lands on OpenRouter: trillion parameters, 49 billion active

Mistral has dropped Large 4 — internally dubbed "Le Chonk" — into public preview on OpenRouter. The new model runs a mixture-of-experts architecture with one trillion total parameters, 49 billion active per forward pass, and it's the company's first natively multimodal design. Context window stretches to one million tokens; max output hits 256 thousand, a serious jump over the previous generation.

Architecture and specs

Choosing MoE with 49 billion active parameters puts Large 4 in the same inference-compute neighborhood as GPT-4o and Claude 3.5 Sonnet, but with a far larger dormant knowledge reservoir. The model accepts text, image, and audio inputs; output is text-only for now. OpenRouter says the API is live immediately, though this is a preview — breaking changes are still possible before general availability.

Pricing: launch discount for two weeks

For the first two weeks OpenRouter is offering 50% off: $0.68 per million input tokens, $2.09 per million output tokens. Cached tokens run $0.07 per million. After the promo the list prices rise to $1.36 and $4.18 respectively — still competitive with comparable closed models, but pricier than open-weights alternatives you can run locally.

Performance claims: "best open-weights model from the US or Europe," no public numbers

The announcement calls Large 4 "the best open-weights model from the US or Europe on aggregated benchmarks" and says it hits state-of-the-art on critical workloads: cyber defense, manufacturing, finance. But the release publishes no concrete scores, names no specific benchmarks (MMLU? GPQA? SWE-bench?), and provides no numerical comparison against Llama 3.1 405B, Nemotron 3 Ultra, or Qwen 2.5 72B. Without public data and independent reproduction, the claims remain a vendor statement, not a measurement.

What this means for developers right now

Immediate availability on OpenRouter lets you test without spinning up your own GPU fleet. The million-token window opens up large-scale RAG, whole-codebase analysis, or long-document processing in a single prompt. The 256k output cap is especially handy for long-form code generation, detailed reports, or full-document translation. On the flip side, preview means potential API changes, latency shifts, and model behavior drift — not something you build critical production on without a fallback.

Bottom line: wait for independent evals

Mistral delivers an impressive spec sheet: large-scale MoE, native multimodality, massive context, attractive launch pricing. But absent detailed benchmarks, third-party evaluations, or a fair fight against existing open-weights rivals, it's too early to say whether Le Chonk actually dethrones the current king. For now: kick the tires, test against your own tasks, and see if those 49B active parameters deliver.