Mixtral 8x22B Instruct: Should You Still Be Running It?

If Mixtral 8x22B Instruct is still in your stack, it is probably there because it worked in 2024 and nobody has had a reason to touch it. That is a fair reason to keep something, right up until the bill is examined. In late 2026 this model is among the most expensive ways to get its level of capability, and its own vendor stopped serving it more than a year ago.

This guide covers what the model is, why it mattered when it shipped, and how to decide whether to move off it.

What it was, and why it mattered

Mistral released Mixtral 8x22B on 17 April 2024 under Apache 2.0, which the announcement called the most permissive open-source licence. It is a sparse mixture-of-experts model with 141B total parameters and 39B active per token, and a 64K context window. The instruct version is Mistral's own fine-tune of the base model.

At the time, that combination was unusual. Mistral claimed it was faster than any dense 70B model and more capable than any other open-weight model, and it reported 90.8% on GSM8K (maj@8) and 44.6% on MATH (maj@4) for the instruct model. It was one of the models that made the case for MoE in open weights: large total capacity, with only a fraction of it running for each token.

It also shipped native function calling, with a constrained output mode on Mistral's platform, and Mistral listed English, French, Italian, German and Spanish as its strong languages. The base model became a foundation for others too; WizardLM-2 8x22B is a well-known fine-tune of it.

What 39B active means against 2026 models

The architecture choice that made Mixtral efficient in 2024 now makes it look heavy. Every token runs through 39B active parameters. Mistral's current Small model, Small 4, runs 6B active out of 119B total and adds image input, a 256K window and optional reasoning.

Active parameters drive serving cost, so the price gap follows. At the time of writing, OpenRouter lists Mixtral 8x22B Instruct at $2 per million input tokens and $6 per million output. Mistral Small 4 is listed at $0.15 and $0.60, by Mistral and on OpenRouter. That is more than thirteen times the input price and ten times the output price, for a model that is two years older.

The memory side cuts the other way, and is worth being honest about. Mixtral's 141B total parameters are not wildly different from Small 4's 119B, so self-hosting either one is a multi-GPU job. Moving does not suddenly make the model fit on smaller hardware; it makes each token cheaper to compute. Dense vs mixture-of-experts lays out that distinction.

Retirement and what it means for you

Mistral's documentation lists Mixtral 8x22B as deprecated on 30 November 2024 and retired on 30 March 2025, with Mistral Small 4 named as the replacement. You cannot call it on Mistral's own platform.

The weights remain on Hugging Face under Apache 2.0, and some third parties continue to host the instruct model. Hosting a retired model is a commercial decision for each provider, which means availability can vanish without the notice a vendor deprecation would give. If you depend on it, either run the weights yourself or keep a tested fallback ready. Model deprecation and migration covers planning for that.

Things that break when you migrate

Moving off an old model is rarely a one-line change. The details that tend to bite with Mixtral:

• Tool-calling format. The model uses Mistral's v3 tokenizer and special tokens such as [TOOL_CALLS] and [AVAILABLE_TOOLS], and the model card says tool call IDs should be exactly nine alphanumeric characters. If you built against the raw format, an OpenAI-style client on a newer model will behave differently.
• Prompt habits. Prompts tuned to coax structure out of a 2024 model often over-specify for a newer one, and can make output worse. Strip them back before judging the replacement.
• Output length. A model with optional reasoning can produce much longer responses if reasoning is left on. Set it explicitly rather than inheriting a default.
• Context assumptions. Code that chunks input to fit 64K is now doing unnecessary work on a 256K model, and may be splitting documents that would read better whole.

A decision rule

Keep Mixtral 8x22B only if one of these is true: you self-host it and the hardware is already paid for, you depend on a fine-tune built on its base, or you need exact reproducibility of outputs produced in 2024.

Otherwise, the vendor's own suggestion is Mistral Small 4, and the price gap alone makes the evaluation worth running. Pull twenty representative requests from your logs, run them through both models on the same prompts, and compare correctness, output length and cost per request. If Small 4 holds up on your tasks — which is likely, but not guaranteed — the migration pays for itself in the first month of any real volume.