Mistral Medium 3.5

An NVIDIA DGX Spark forum post benchmarks the NVFP4-quantized Mistral Medium 3.5 128B with the EAGLE draft model on a single GB10 node using vLLM 0.20.2rc1. Reported throughput at 16k context ranges from roughly 6.9 tok/s for short Q&A to about 9.1 tok/s for a 2048-token code generation, with vLLM showing ~72.9 GiB of target-model memory. The run uses --load-format auto for stability, since --load-format fastsafetensors triggered OOM crashes on the 128B weights at 16k context on this hardware. The author also measured prompt-processing throughput via TTFT and notes that a 3026-token warmup prompt hit 166.2 tok/s, indicating practical single-node serving viability for the NVFP4 build.

Towards AI's hands-on review tests Mistral Medium 3.5 across 18 coding tasks at multiple reasoning-effort settings, comparing it against the retired Magistral reasoning model and Devstral 2. The author finds the unified 128B weights match or beat the specialized predecessors on coding and agentic tool use. Released April 29, 2026, Mistral Medium 3.5 is described as a dense 128B open-weight model with a 256K context, configurable reasoning effort, and Modified MIT licensing. The piece positions the release as a counter-trend to labs shipping separate reasoning and base models, consolidating those roles into one set of weights.

Mistral announced Mistral Medium 3.5 on May 22, 2026 as a 128B dense open-weight flagship that merges instruction-following, reasoning, and coding into a single set of weights with a 256k context window. The model runs self-hosted on as few as four GPUs and powers new Mistral Vibe remote agents plus a Work mode in Le Chat for multi-step tasks like research, analysis, and cross-tool workflows. Pricing is set at $1.5 per million input tokens and $7.5 per million output tokens, with availability on Pro, Team, and Enterprise plans. Reasoning effort is configurable per request, letting the same weights handle both quick answers and longer analytical workloads. Coding sessions can be spawned from the Mistral Vibe CLI or Le Chat, run asynchronously in the cloud, and a local CLI session can be teleported up to the remote runtime. The release marks Mistral's first flagship merged model in public preview, positioned as the default engine inside both Vibe and Le Chat.

Mistral AI documents Mistral Medium 3.5 as a frontier-class multimodal model optimized for agentic and coding use cases, released as open weights under a Modified MIT license and published April 28, 2026. The page confirms a 256k context window with pricing of $1.5 per million input tokens and $7.5 per million output tokens. Supported endpoints include /v1/chat/completions, /v1/conversations, /v1/agents, and /v1/batch, with features such as structured outputs, function calling, predicted outputs, document QnA, prefix caching, and batching clearly listed. The documentation entry positions Mistral Medium 3.5 as a multimodal model with broad API surface area, including built-in tools and agent/conversation routes. It is listed alongside sibling Mistral offerings such as Z.ai GLM 5.3, Z.ai GLM 5.2, and Shieldstral 1.0. The page serves as the canonical spec anchor for developers integrating the model via Mistral's own APIs, independent of any third-party gateway.

VRLA Tech's April 29, 2026 coverage details Mistral Medium 3.5's self-hosting profile: 128B dense parameters, 256k context, Modified MIT license, and deployable on as few as four GPUs, with availability on Hugging Face, Le Chat, Mistral Vibe, and NVIDIA NIM. The article reports a 77.6% score on SWE Bench Verified for coding tasks. API pricing is listed at $1.50 per 1M input tokens and $7.50 per 1M output tokens, with vision support via a custom-trained encoder and configurable per-request reasoning. The piece frames the release as lowering the hardware bar for running flagship-class open-weight models on a single multi-GPU server rather than hyperscaler infrastructure.

NVIDIA published an NVFP4-quantized build of Mistral Medium 3.5 128B on Hugging Face, optimized with NVIDIA Model Optimizer for deployment on NVIDIA hardware. The card reiterates the 128B dense architecture, 256k context, multimodal text-and-image input, and support for function calling, JSON output, and configurable reasoning effort. Third-party terms apply: NVIDIA disclaims authorship, and use is governed by Mistral's Modified MIT license. The quantization enables lower-memory serving of the 128B weights while preserving the underlying Mistral Medium 3.5 capabilities for production inference on NVIDIA platforms.

Mistral released Mistral Medium 3.5 as a dense 128B parameter model with a 256k context window, unifying instruction-following, reasoning, and coding in one set of weights. The official model card on Hugging Face confirms it replaces Mistral Medium 3.1 and Magistral in Le Chat, and Devstral 2 in the Vibe coding agent. Key capabilities listed on the card include configurable per-request reasoning effort, native function calling, JSON output, multilingual support across major languages, and a custom vision encoder trained for variable image sizes. The model is released under a Modified MIT License for commercial and non-commercial use, with an EAGLE speculative-decoding companion available for vLLM and SGLang inference.

Mistral's official documentation changelog records the release of Mistral Medium 3.5 (mistral-medium-3-5) in the April 2026 section, explicitly naming the model id and placing it among Mistral's own model release entries. This is a first-party confirmation from the model author that Mistral Medium 3.5 exists as a disti Beyond confirming the model's existence and release, this changelog entry is significant because it sits alongside other model and API release lines for Mistral, establishing that Mistral Medium 3.5 was shipped as a discrete versioned model update rather than a silent default-alias change. Readers tracking the Mistral