Mistral Medium 3.5

An NVIDIA developer-forums performance report from May 4, 2026 documents running the 128B Mistral Medium 3.5 weights in NVFP4 quantization alongside Mistral's official EAGLE draft model on a single DGX Spark / GB10 node using vLLM. The test reported target model memory of 72.91 GiB and a KV cache supporting roughly 89,500 tokens at 16k context. Generation benchmarks using temperature 0.0 measured 6.9 tok/s on a 256-token Q&A, 8.8–8.9 tok/s on 512–701 token code and JSON outputs, and 9.1 tok/s on a 2048-token long-code task. The report notes that the --load-format fastsafetensors loader caused OOM instability at 16k context while --load-format auto was stable but slower.

Mistral Medium 3.5 128B is described as Mistral AI's first flagship merged model, a dense 128B-parameter transformer with a 256K-token context window that unifies instruction-following, reasoning, and coding in a single set of weights. Per NVIDIA NIM documentation, it replaces Mistral Medium 3.1 and Magistral in Le Chat, and also replaces Devstral 2 in the Mistral Vibe coding agent. The model was released on April 29, 2026 and supports configurable per-request reasoning effort, letting the same weights handle quick chat or complex agentic runs. Mistral trained the vision encoder from scratch for variable image sizes and aspect ratios, enabling multimodal text-plus-image input alongside function calling and JSON output. The release is open-weight under a Modified MIT license for commercial and non-commercial use.

Think Facility's model profile corroborates Mistral Medium 3.5's April 28, 2026 release from Mistral's changelog, with API model ID mistral-medium-3-5, 256K context, and pricing at $1.50 input and $7.50 output per million tokens. The page contextualizes the model's pricing and context window against other recent frontier and Mistral models. The profile notes Medium 3.5 consolidates Mistral's prior Magistral reasoning line and Devstral coding line into a single frontier model. It is described as a French-built model released under a modified MIT license, currently live and available for use.

Mistral's Hugging Face model card details Mistral Medium 3.5 as a dense 128B-parameter model with 256k context, handling instruction-following, reasoning, and coding in a single set of weights. The model accepts text and image input, producing text output, with configurable reasoning effort per request toggling between fast replies and deeper agentic runs. The card notes Medium 3.5 replaces Mistral Medium 3.1 and Magistral in Le Chat, and Devstral 2 in the Vibe coding agent. Mistral also released an EAGLE speculative-decoding draft model for vLLM and SGLang inference acceleration, and issued a Transformers config fix correcting a long-context performance bug that had affected earlier GGUFs.

Mistral released Mistral Medium 3.5 (v26.04) on April 28, 2026, describing it as a frontier-class multimodal model optimized for agentic and coding workloads, released as open weights under a Modified MIT license. The model offers a 256k context window with published pricing of $1.50 per million input tokens and $7.50 per million output tokens. The official documentation lists supported features including structured outputs, function calling, predicted outputs, document QnA, prefix caching, and batching across endpoints /v1/chat/completions, /v1/conversations, /v1/agents/v1/conversations, and /v1/batch. These capabilities position Medium 3.5 as a unified frontier model for Mistral's platform.

Mistral Medium 3.5 is tracked on BenchmarkList with pricing of $1.50 per 1M input tokens and $7.50 per 1M output tokens, benchmarked against frontier peers including Claude Fable 5.1, Claude Opus 5, Kimi K3, Qwen3.8 Flash, Qwen3.8-2.4T-A95B and GLM 5.3. The listing compiles 45 evaluation rows across reasoning, coding, Recent head-to-head numbers place Mistral Medium 3.5 at 94.2% on Tau2-Bench Telecom (June 10, 2026), 33.3% on Terminal-Bench Hard (June 10, 2026), Elo 936 on GDPval-AA (September 2, 2026), and 15.1% pass@1 on Tau3-Banking (September 2, 2026), with Fable 5 leading most agentic indices at 98.5% and 62.9% respectively, an