Mistral AI just killed half its model family. Here is what survived and why.

Mistral AI's model catalog in mid-2026 offers a rare window into how an AI company's strategy evolves from its experimental phase to production discipline. The portfolio, which includes 36 active models across four tiers (Featured, Frontier, Other, and Legacy), shows a company that has made hard choices about which experiments graduate and which get retired.

The most telling signal is the deprecation table. Nineteen models are listed with retirement dates between May and August 2026, including every model from the 2024-2025 era that defined Mistral's early reputation: Mistral Large 2.0 and 2.1, Pixtral Large, Codestral 24.05, Mistral Small 1.0 through 3.2, and the entire Magistrate line. The message is unambiguous: the models that put Mistral on the map are being cleared out to make room for a new generation that integrates lessons learned across every product line. Mistral AI just raised €600 million. The pressure to…

The two-tier system

Mistral now operates a clear two-tier model strategy. The open tier includes Mistral Medium 3.5, Mistral Small 4, Ministral 3 variants down to 3B, Voxtral TTS, and Leanstral 1.5, with weights and self-hosted deployment options available. The premier tier, which includes OCR 4, OCR 3, Voxtral Mini Transcribe 2, Codestral, Mistral Moderation 2, and Embed models, is API-only, typically with higher performance for specific verticals but no weight access.

This split reflects a strategic judgment that not all AI capabilities should be open-weight. OCR, speech transcription, and moderation handle data-sensitive workloads where enterprises need a governed API even if they could theoretically self-host. The premier tier also generates the predictable revenue stream that funds the open models. Two AI labs just proved why open models win in…

The small-model thesis

The Featured Models list, which includes Mistral Medium 3.5, Mistral Small 4, Voxtral Mini Transcribe Realtime, Voxtral TTS, OCR 4, and Devstral 2, reveals the company's current product thesis: small, specialized models that outperform their size class through architectural innovation rather than parameter count. MiniMax's M3 just wrote its own CUDA kernel, and opened…

Mistral Small 4 is described as a hybrid model unifying instruct, reasoning, and coding in a single efficient model, an attempt to collapse three capability axes that competitors typically split across separate model variants. Devstral 2 is positioned as a frontier code agents model for solving software engineering tasks, directly targeting GitHub Copilot and Cursor's agentic workflows. OCR 4 adds bounding boxes and block classification to document extraction, turning a commodity text recognition tool into a structured data pipeline component.

None of these models exceed what Mistral classifies as medium in its naming convention. The company is betting that the future of production AI belongs to models that can run on a single GPU, not the multi-node clusters required for 400B-parameter frontier models. Cursor just turned your iPhone into a serious coding machine

What got deprecated and why

The deprecation list offers clues about which experiments failed to find product-market fit. The Magistrate line, medium and small variants, is retired wholesale, with the final versions from September 2025 receiving barely eight months of support. Magistrate was Mistral's attempt at a reasoning-specialized model line, analogous to OpenAI's o-series, but it appears to have been subsumed into the unified Mistral Small 4 and Medium 3.5 architectures.

Pixtral Large, the company's flagship vision-language model from late 2024, is also on the retirement track. Its recommended alternative is Mistral Medium 3.5, suggesting that the company's latest generalist multimodal model now exceeds the performance of its dedicated vision model, a common convergence pattern as vision-language capabilities become standard in base models rather than separate products. ViQ just gave multimodal AI the one thing it needed:…

Devstral Small 1.0 and 1.1 are deprecated in favor of Mistral Medium 3.5, implying that even Mistral's own coding specialist has been overtaken by the generalist frontier model. But Devstral 2 remains a featured product, suggesting a new architecture that earns its keep as a standalone specialist.

Strategic vectors for the second half of 2026

Three patterns stand out for the coming months. First, the Voxtral line, TTS and Transcribe, represents Mistral's push into speech modalities, a domain where the company competes with ElevenLabs, OpenAI's Whisper, and Google's Chirp. The Realtime variant of Voxtral Mini Transcribe signals a low-latency use case that could power live captioning or voice agent pipelines. The missing 'ums' and 'uhs' that finally make AI speech…

Second, the continued support for Ministral 3 variants down to 3B shows a commitment to edge deployment. At 3 billion parameters, a model can run on phones and IoT devices, opening use cases that 70B+ models cannot address.

Third, Leanstral 1.5, an updated code agent for Lean 4 formal proof engineering, remains in labs rather than premier or open, indicating that automated theorem proving is still a research project rather than a product. But its survival through a round of heavy deprecation suggests Mistral sees long-term value in formal methods, possibly as infrastructure for model reliability guarantees. Mistral buys its way into physics, and a race against…

The mid-2026 Mistral portfolio is not the same company that shipped Mistral 7B in 2023. It is leaner, more disciplined about which bets to keep, and organized around production deployment economics rather than benchmark chasing. The deprecation list is the most honest signal a model vendor can send about what it believes in, and what it has stopped believing in.