Mistral: Mistral Medium 3.5

Mistral AI

•

Mistral Medium 3.5

•

Released

April 2026

Mistral AI
Mistral Medium 3.5

Dense 128B open-weight model merging chat, reasoning and coding, with per-request reasoning effort and 256K context.

Modality:

Text

Image

PDF

model ID

mistralai/mistral-medium-3-5

Start for free

View API Docs

Output Speed *

159.52

tok/s

Intelligence Index *

14.2

/ 100

Context Window *

262144

tokens

Input price

Anytoken

Output price

Anytoken

One 128B Model for Chat, Reasoning and Coding Agents Mistral Medium 3.5 is Mistral AI's first flagship merged model: a dense 128B-parameter model that combines instruction-following, reasoning and coding in one set of open weights. It replaces Medium 3.1 and Magistral in Le Chat and Devstral 2 in Mistral's Vibe coding agent. It accepts text and image input, has a 256K-token context window, and makes reasoning effort a per-request setting. It suits coding agents, tool-using workflows and document analysis where one model must handle both quick replies and long reasoning runs. Access Mistral Medium 3.5 via AnyAPI.ai

Performance

Agentic Coding Results From a General-Purpose Dense Model

Mistral Medium 3.5 merges instruction-following, reasoning and coding into one 128B dense model. Its strongest evidence is agentic coding: Mistral reports 77.6% on SWE-Bench Verified, ahead of Devstral 2 at 72.2% and Qwen3.5 397B A17B. These are vendor-reported figures, but they suggest a general model can now cover coding-agent work without a separate specialist checkpoint. In production, one endpoint can serve fast replies with reasoning_effort set to none and long agent runs with high, though reasoning mode adds latency.

Benchmarks

Independent Measurements: Above-Median Intelligence, Good Throughput, Slow First Answer

Artificial Analysis scores Mistral Medium 3.5 at 14 on its Intelligence Index v4.3.2, above the median of 8 for comparable open-weight models. That is solid, not frontier-level. Output speed is about 168 tokens per second on Mistral's API, faster than the 128 tokens per second average. IFBench instruction following is 68.8%. The same provider page lists about 14 seconds to the first answer token on a 10K-input workload, so responsiveness depends heavily on reasoning effort.

Output Speed

159.52

tok/s

Intelligence Index

14.2

/ 100

MMLU *

Broad world knowledge and problem-solving

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.

HLE *

Adherence to multi-step structured instructions.

LiveCodeBench *

Tool-calling reliability in long agentic loops.

Technical Specifications

What the model supports

Mistral Medium 3.5 is a dense 128B model that takes text and image input and returns text only. Its 256K-token context window is a separate limit from output length, and a maximum output value is not consistently documented across hosts, so budget completions explicitly. Reasoning is set per request (none or high), which matters most in production: the same deployment can serve cheap extraction and expensive agent runs. Open weights ship under a modified MIT license.

Verified Specifications —

Mistral Medium 3.5

Input modalities

Text

Image

PDF

output modalities

Text

Context window

262144

tokens

Maximum output tokens

262144

Reasoning

Yes

Knowledge cutoff

April 2026

Pricing (standard)

AnyTokens in

AnyTokens out

Limitations & Trade-offs

Where Mistral Medium 3.5 falls short

Reasoning mode adds noticeable latency. Artificial Analysis lists about 14 seconds to the first answer token on a 10K-input workload, even though output speed is about 168 tokens per second. Interactive chat, autocomplete and voice-style UIs should use reasoning_effort none or a smaller, faster model such as Mistral Small 4. Reserve high effort for background agent runs and batch-style analysis.

Factuality is a weak spot. Third-party trackers report an AA-Omniscience index of about -36, which means wrong answers outnumber correct ones on that knowledge test. Artificial Analysis also places its Intelligence Index at 14, solid for an open-weight model but not frontier-level. For knowledge-heavy Q&A or RAG, ground answers in retrieved sources and consider a stronger model for high-stakes factual work.

It sits in a premium cost tier. Artificial Analysis labels both input and output pricing as expensive against comparable open-weight models, and the dense design activates all 128B parameters on every token. Long reasoning outputs and agent loops compound this. For high-volume classification, extraction or simple chat, a smaller Mistral model is usually the more economical default.

Preview status and host-dependent limits. Mistral described the model as being in public preview, and Azure AI Foundry lists a preview lifecycle with 128K context and a 4,096-token output cap, below the 256K model-card figure. Output is text only, with no image or audio generation. Verify the context and output limits of your specific route before designing long-document or long-generation workflows.

Best-Fit Workloads

Where this model earns its place

Coding agents

Mistral reports 77.6% on SWE-Bench Verified, and Medium 3.5 is the default model in Mistral's Vibe coding agent. That fits multi-file edits, repository-scale context within 256K tokens, and tool-driven fix loops. Use high reasoning effort for hard tasks. The score is vendor-reported, so run your own repository tests, and watch cost on long agent loops.

Tool-using research and workflow agents

The model supports native function calling in both modes and is reported at 94.2% on the τ²-Bench Telecom multi-turn tool-use benchmark. That makes it a candidate for agents that triage tickets, query systems and synthesize research. This is a narrow benchmark; broader agentic suites rank it lower, so test on your own tool schemas before trusting long autonomous runs.

Document and screenshot analysis

Medium 3.5 accepts text and image input through a vision encoder trained from scratch for variable image sizes and aspect ratios, and Azure lists PDF input. Artificial Analysis reports 69.3% on its AA-LCR long-context reasoning test. It suits invoice parsing, chart reading, UI screenshots and report extraction. Output is text only, so structured extraction should use JSON output.

Tiered chat and reasoning assistants

Because reasoning effort is a request parameter, one model can answer routine questions with reasoning off and escalate complex ones to high. This replaces maintaining separate chat and reasoning models, as Mistral did with Medium 3.1 and Magistral. The trade-off is that high effort increases latency and output tokens, so route by task difficulty.

Pricing in anytokens via AnyAPI

Input

₳

Output

₳

Cache write

—

₳

Cache read

—

₳

Get API access

Read Docs

Integration

Access Mistral Medium 3.5 via AnyAPI.ai

Access Mistral Medium 3.5 through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Mistral Medium 3.5 without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

One API integration

Access Mistral Medium 3.5 and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

Easy model switching

Test Mistral Medium 3.5 against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

Flexible for production

Use Mistral Medium 3.5 from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

Multi-model applications

Use Mistral Medium 3.5 for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free

Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

What is the Mistral Medium 3.5 context window?

Mistral Medium 3.5 has a 256K-token context window per Mistral's model card, listed as 262,144 tokens by OpenRouter and Artificial Analysis. It accepts text and image input and returns text only. Context is separate from maximum output, and some hosts apply lower limits, such as 128K on Azure AI Foundry.

How does reasoning effort work in Mistral Medium 3.5?

Mistral Medium 3.5 uses a per-request reasoning_effort setting. Mistral recommends "none" for fast replies, chat and extraction, and "high" for complex prompts, coding, research, math and agentic work. For high effort it suggests temperature 0.7. One model therefore replaces separate instruct and reasoning variants.

How good is Mistral Medium 3.5 for coding?

Mistral reports 77.6% on SWE-Bench Verified for Mistral Medium 3.5, above Devstral 2 at 72.2%. These are vendor-reported figures. Artificial Analysis gives a general Intelligence Index of 14, above the median for comparable open-weight models. Test on your own codebase before replacing an existing coding model.

Is Mistral Medium 3.5 open-weight?

Yes. Mistral released Mistral Medium 3.5 as open weights on Hugging Face under a modified MIT license. Mistral says it can be self-hosted on as few as four GPUs, depending on the deployment. The license includes conditions for large-revenue companies, so review the terms before commercial self-hosting.

Is the Mistral Medium 3.5 API expensive?

Mistral Medium 3.5 sits in a premium cost tier. Artificial Analysis labels both its input and output pricing as expensive compared with similar open-weight models. It is better suited to agents, coding and complex analysis than to high-volume simple tasks. Current customer-facing pricing is shown in the AnyTokens pricing component on this page.

* Benchmark data source: Artificial Analysis artificialanalysis.ai