Ministral 3 8B: The Small Mistral Model to Start With

If you need a small model and do not yet know how small, Ministral 3 8B is the sensible place to start. It is large enough to hold a conversation and follow long instructions, small enough to run on a single consumer GPU, and cheap enough on an API that the bill rarely becomes the deciding factor.

Mistral released it on 2 December 2025 as the middle size of the Ministral 3 family, between the 3B and the 14B, all under Apache 2.0. This guide covers what it is, how its pricing behaves, and how to tell whether you should move up or down a size.

The model in brief

The model card lists an 8.4B-parameter language model plus a 0.4B vision encoder. Mistral's paper on the family describes it as dense; it accepts text and images, and returns text. Mistral's documentation gives a 256K-token context window, function calling, structured outputs and document question answering.

Like its siblings, it was produced by cascade distillation, Mistral's term for repeatedly pruning a larger model and continuing training with distillation, and it ships in base, instruct and reasoning variants. Mistral's suggested uses are telling: a local daily-driver assistant, chat in constrained environments, image and document description, translation, and specialised agentic tasks.

Flat pricing changes the arithmetic

Most hosted models charge several times more for output than input. Ministral 3 8B does not. Mistral lists it at $0.15 per million tokens for both, and OpenRouter shows the same $0.15 input and $0.15 output at the time of writing.

That matters for workloads that generate more than they read. Summarising a short document into a long report, drafting replies, or translating text all produce output roughly the size of the input or larger. On a model with a five-to-one output premium, those jobs cost far more than the input price suggests. Here they do not.

It also makes cost estimation trivial. Count total tokens, multiply by one rate. Input vs output token pricing covers why that asymmetry usually decides the bill, and why a flat rate removes a whole class of surprises.

The batch endpoint

The catalogue also lists ministral-8b-2512:batch. It is the same model behind an asynchronous endpoint: you submit a file of requests, results come back later, and the price halves, to $0.075 per million tokens each way on OpenRouter at the time of writing.

If the work does not need an answer within seconds — nightly tagging, backfilling descriptions for an image archive, translating a content catalogue — there is no reason to pay the interactive rate. Batch API savings walks through when the latency trade is worth it.

Where the 8B sits in its family

Mistral publishes instruct scores for all three sizes, which makes the trade unusually easy to see:

• Arena Hard — 3B 0.305, 8B 0.509, 14B 0.551
• WildBench — 3B 56.8, 8B 66.8, 14B 68.5
• MATH Maj@1 — 3B 0.830, 8B 0.876, 14B 0.904
• MM MTBench — 3B 7.83, 8B 8.08, 14B 8.49

The jump from 3B to 8B is large on the open-ended tests: Arena Hard rises by about two-thirds. The jump from 8B to 14B is small by comparison. That shape is the argument for the 8B as the default. It captures most of the family's conversational quality at a little over half the 14B's size.

These are vendor-run numbers, and Arena Hard and WildBench reward chat behaviour rather than your specific task. Treat them as a guide to the relative gaps, not as a forecast of your results.

What 12GB of VRAM buys you

The model card says the FP8 weights run in about 12GB of VRAM, less if you quantise further. That is a single mid-range consumer GPU or a well-specified laptop, which is why Mistral pitches it as a local daily driver.

Running locally changes the economics entirely: no per-token cost, no data leaving the machine, and latency set by your hardware rather than a shared queue. The trade is that you own the serving stack: updates, monitoring, and the hardware sitting idle when nobody is using it.

One inconsistency in the model card is worth knowing. It recommends a temperature below 0.1 for production work, yet its own example scripts use 0.15. If you copy the example and see drifting answers on a repetitive task, the temperature is the first thing to lower. The vLLM flags needed for tool calling are the same as for the 3B; the Ministral 3 3B guide has the full serve command.

Mistral's launch post also claims the Ministral 3 instruct models often produce an order of magnitude fewer tokens than comparable models. That is worth checking rather than accepting: shorter answers lower cost and latency, but only if they are still complete.

When to move up or down a size

• Move down to 3B when the task has a fixed output shape — a label, a schema, a short translation — and you need the smallest footprint or the lowest latency.
• Stay on 8B for chat, document description, multi-paragraph instructions and light agent work with a small tool set.
• Move up to 14B when you have the memory and the 8B visibly misreads complex instructions or fumbles multi-step tool calls. The Ministral 3 14B guide covers it.

The way to decide is to measure, not to read tables. Take fifty real tasks, write down what a correct answer looks like, and run them through the 8B and its neighbours at the same temperature. Count failures and total output tokens. If the 8B fails on fewer than you can tolerate, it is your default, and you can route the hard cases upward rather than paying for the larger model on every call.