Mistral 3, Large 3 & Vibe: Why the Latest Update Puts Europe's AI Hope Back in the Game

Quick reality check: where Mistral stood six months ago

In late 2025 the mood around Mistral was sober. Llama 4, DeepSeek-R1, Qwen3.5 and Kimi K2.5 had carved up the open-weights market. Mistral Large 2 was solid but no longer a frontier model. Le Chat was a decent ChatGPT competitor – without a wow moment.

If you had asked in spring 2026 whether Europe still had its own frontier lab, you would have gotten a cautious "kind of."

Six months later it looks different. The reason is a remarkably dense release cadence out of Paris.

Mistral 3: the December 2025 breakout

On December 2, 2025, Mistral announced Mistral 3 – shipping four models at once:

Three things matter here:

1. Large 3 is Mistral's first Mixture-of-Experts model since Mixtral – and at 675B total parameters it sits right next to DeepSeek and Kimi.
2. Everything is Apache 2.0 – truly free open weights, not "modified MIT with MAU thresholds" like Llama or Kimi.
3. Multimodal and multilingual out of the box – 40+ languages natively, including vision. Exactly where US models often stumble on German, French or Italian content.

Large 3 debuted as #2 in the OSS Non-Reasoning category on LMArena – a comeback nobody priced in back in November. A reasoning variant was announced in parallel.

Why it's technically interesting

Mistral co-optimized Large 3 with vLLM, Red Hat and NVIDIA. Concretely:

• NVFP4 checkpoint out of the box – runs on a single 8×H100 or 8×A100 node via vLLM
• TensorRT-LLM and SGLang support at launch (not months later)
• Blackwell-optimized MoE and attention kernels for GB200 NVL72

For builders: you can self-host Large 3 from day one without waiting for community quantizations. For GDPR-pressured organizations that's gold. (More on quantization formats →)

Medium 3.5: the new default workhorse (May 2026)

In May 2026 came the update everyone is talking about: Mistral Medium 3.5.

The trick: Medium 3.5 fuses three things that used to be separate models – instruction following, reasoning and coding. You pick per request how deeply the model "thinks." Same idea as GPT-5.2 and Claude Opus 4.5, but in a 128B model you can host yourself.

Real-world performance: Medium 3.5 beats the old Large-2 generation on most coding and agentic benchmarks and gets close to frontier models – at a fraction of the inference cost.

For our stack: Medium 3.5 is a realistic candidate as the default model for tool-calling agents without handing control to OpenAI or Anthropic.

Le Chat becomes Vibe: the move that's actually new

On May 22, 2026 Mistral launched Vibe – and soon after rebranded Le Chat entirely into Vibe. This is more than a rename.

What Vibe brings technically:

1. Remote coding agents in the cloud. Start a coding task in the Vibe CLI or chat – the job keeps running in a cloud sandbox even after you close your laptop.
2. Teleport. Lift a local CLI session into the cloud. The workflow many miss in Cursor and Claude Code.
3. Work Mode in Le Chat / Vibe. Multi-step tasks with parallel tool use – classic agentic AI, inside the Mistral stack, no MCP duct-tape.
4. VS Code extension. Vibe ships as a plugin – competing directly with Cursor, Composer 2 and Claude Code in the IDE.

This is exactly the move that pulls Mistral out of the pure model-provider role and into the tool layer – where the money is being made right now (see the Cursor/Kimi discussion →).

Where Mistral sits in the stack now

Cross-referencing our current open-source LLM comparison, the picture is clear:

That last point isn't trivial. Mistral is the only frontier lab headquartered in the EU. Anyone rolling out AI in regulated sectors (healthcare, public sector, banking) in 2026 should at least have evaluated whether the Mistral line is the politically and legally cleaner path – even if US models still lead on isolated benchmarks.

What we're actually doing

At Till Freitag we're running three parallel experiments:

1. Medium 3.5 as backend for our internal agents. Comparing latency and tool-calling accuracy against Claude Sonnet 4.5 on real monday.com workflows.
2. Vibe in the toolbox rotation. Our vibe coders test Vibe alongside Cursor and Claude Code – focus on remote sessions for longer refactors.
3. Large 3 for multilingual content pipelines. Especially DE/FR/IT content where English-centric models often miss the tone.

Early results are honestly better than I expected. Medium 3.5 in particular holds up surprisingly well in our RevOps use cases.

What to do now

1. If you're on Llama or Mixtral today: Test Medium 3.5 – it's the direct upgrade path without an architecture rewrite.
2. If you use Cursor or Claude Code: Try Vibe for a week. The remote sessions especially are a different workflow.
3. If you're building a self-hosting stack: Large 3 on vLLM is the cleanest open-weights path right now with a real Apache 2.0 license.
4. If you handle regulated data in the EU: Mistral belongs on the shortlist from now on.

Bottom line

Mistral was filed under "had its moment" in early 2026. With Mistral 3, Medium 3.5 and Vibe the lab is back at the frontier in six months – not on every single benchmark, but in the overall package of open licenses, EU hosting, honestly open models and a serious tool layer.

The latest update isn't just "pretty good." It's the most important sign of life a European AI lab has shown in years.

---

→ Open-source LLM comparison: 20+ models → Kimi K2.5 & Cursor: what the Composer 2 move means → GGUF, GGML, Safetensors – formats for local AI → Our AI services at Till Freitag