mistral-medium-3.5-128b

Overview. Mistral Medium 3.5 is a dense 128B-parameter model from Mistral AI, released as open weights under a Modified MIT licence. It accepts text and image input, produces text output, and has a 262,144-token context window. Developers can reach for it when they want one open-weight model that handles multimodal input and can switch extended reasoning on or off.

Key strengths. The model card documents a reasoning_effort setting with 'none' and 'high' values. This lets one model serve both direct instruction-following and longer reasoning traces. The architecture is dense rather than mixture-of-experts, so all 128B parameters are active on every token. The 262,144-token context window is confirmed by the upstream configuration. No benchmark scores such as MMLU or Arena were captured for this snapshot, so we make no benchmark-based quality claims here.

When to use it. Use this model for long-document and multi-turn work that needs a large context window, especially when image input is part of the workload. Turn reasoning on for harder, multi-step problems and off for simpler requests where you want a direct answer. It suits interactive chat-style Inference, but latency has not been measured, so test it against your own response-time requirements.

Limitations. No maximum output length or training-data cutoff has been verified, so don't assume either. The model takes text and image input only and produces text only, with no audio in or out. The Modified MIT licence is not standard MIT. It restricts use by organisations above a stated revenue threshold unless they obtain a separate commercial licence from Mistral AI, so check the licence text before deploying. Mistral also warns that older Transformers configurations, and GGUF files generated from them, degrade at long context, so use the corrected configuration when self-hosting.

Sovereignty note. The weights are openly published, so you can run this model on infrastructure you control. That is the clearest route to Sovereign deployment. We have not verified the creator's CLOUD Act exposure, any training-on-input policy, or endpoint data residency for this model, so check the specific endpoint before treating it as Sovereign.