What changed in AI today

• Four sizes, 5M to 209M parameters; the 209M model ranks first on GIFT-Eval, TIME and fev-bench.
• Outperforms the previous state of the art with two-thirds of its parameters, per Mistral's own benchmark results.
• Runs without a GPU: 10 ms with the 5M model, 140 ms with the 209M model on a standard server CPU.
• Available through Mistral Forge for fine-tuning and on-premise or air-gapped deployment; GitHub repository published.

Why it matters for marketers: Content and analytics teams can forecast from sensor or business data on CPUs, and fine-tune on proprietary data via Forge, without building per-dataset models.

• Preview API is live on Mistral Studio; open weights are promised by the end of the month.
• Model has 1 trillion parameters, 52 billion active, and is natively multimodal, per Mistral.
• Scores 82% on a real-vulnerability reproduce-and-patch test and 93% on Cybench, which Mistral says is a top open-weight result.
• Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters.

Why it matters for marketers: Content and security teams gain an open-weight option with fewer provider-level refusals, but should wait for released weights and independent checks before relying on the self-reported benchmarks.