December 2, 2025
Benchmarks, broken links & EU bravado
Mistral 3 family of models released
Open weights drop, Europe cheers, devs ask: “How does it stack up”
TLDR: Mistral launched the Mistral 3 lineup—open models from small to a new “Large 3”—under a permissive license and tuned with NVIDIA for easy deployment. Commenters applauded the openness and Europe’s push, but clamored for clear head‑to‑head comparisons, asked about tool support, and griped about a broken download link.
Mistral just tossed a shiny new toy into the AI arena: Mistral 3—a family of open models ranging from smaller “Ministral” versions (3B, 8B, 14B) to Mistral Large 3, a beefy “mixture‑of‑experts” system (think several specialists working together) released under the Apache 2.0 open license. They teamed with NVIDIA, vLLM, and Red Hat to make it fast and cheap to run, and it’s already flexing on the LMArena leaderboard at #2 among open non‑reasoning models.
But the real show is the comments. Fans are hyped yet hungry for receipts: timpera cheered the release but begged for simple, side‑by‑side comparisons with OpenAI, Google, and Anthropic so everyone can see where Mistral stands. Builders like codybontecou immediately asked the only question that matters for apps: do these models support tool use and structured output (aka reliably calling APIs and returning tidy JSON)? Meanwhile, the skeptics rolled in hot: simgt questioned the business logic of releasing great open weights at all, hinting that benchmark games and PR stunts rule the day.
Amid the hype and theorycrafting, one very real drama popped up: broken links. hnuser123456 flagged a bumpy Hugging Face rollout and posted direct links to the models. And yes—Europe took a victory lap, with yvoschaap upvoting Mistral as the continent’s boldest AI play. Open models, leaderboard clout, and a little chaos? The internet’s favorite combo.
Key Points
- •Mistral released the Mistral 3 family: Ministral 3 (14B, 8B, 3B) and Mistral Large 3, all under Apache 2.0.
- •Mistral Large 3 is a sparse MoE model with 41B active and 675B total parameters, trained on 3000 NVIDIA H200 GPUs.
- •After post‑training, Large 3 achieves parity with leading instruction‑tuned open‑weight models, with image understanding and strong multilingual performance.
- •An NVFP4 checkpoint built with llm‑compressor enables efficient serving via vLLM on Blackwell NVL72 and single 8×A100/8×H100 nodes; NVIDIA enabled TensorRT‑LLM and SGLang support.
- •NVIDIA integrated Blackwell attention and MoE kernels, added disaggregated prefill/decode and speculative decoding; edge deployments are optimized on DGX Spark, RTX PCs/laptops, and Jetson devices.