Mistral 3.5
#106 in Frontier-Spraakmodellemistral · v3.5 · siet 2026-04-29 · 12× · tolest 30. Juni 2026
12
Momentum
Mistral Medium 3.5 is Mistral's first "flagship merged model": a dense (non-MoE) language model with 128 billion parameters that unifies instruction-following, reasoning, and coding in a single set of weights. It features a 256,000-token context window, a natively trained vision encoder (text and image input), and configurable reasoning effort per request. The model is available as open-weight under a modified MIT license and supersedes its predecessors Mistral Medium 3.1, Magistral, and Devstral 2.
Momentum-Verloop
04.04.03.07.
Features
| Key Benchmark (%) | 77.6% on SWE-Bench Verified; 91.4% on τ³-Telecom |
| Context Window (Tokens) | 256,000 tokens (262,144) |
| License | Open weights, modified MIT license |
| Multimodality | Text and image input, text output (proprietary vision encoder trained from scratch) |
| Platform | Le Chat, Mistral API, Mistral Vibe, Hugging Face, NVIDIA NIM/build.nvidia.com, vLLM/SGLang/Ollama (self-hosting) |
| Price per 1M Tokens | $1.50 (input) / $7.50 (output) |
| Release Date | April 29, 2026 |