Mistral Small 4 First Look: The Free Model That Keeps Your Data in the Building
The gate: Five stars need 21+ and Australian Fit of at least 4/5.
Skip to the verdictLaunched: 16 March 2026 · Reviewed: 17 March 2026 · Score: 18/25 ⭐⭐⭐⭐ Very Good
Verdict in one line: The strongest "your data never leaves Australia" option of early 2026 — if you have the hardware or a hosting partner to run it.
⚠️ STALENESS FLAG — 20 September 2026. This first look was published 17 March 2026 and is 187 days old. It has not been re-verified against the vendor since. AI products move fast: pricing, model versions, plan names and Australian availability may all have changed. Read the verdict below as a snapshot of March 2026, not as a current recommendation. (House rule: any review over six months old carries this flag until it is updated.)
What launched
On 16 March 2026 French lab Mistral AI released Mistral Small 4 under the Apache 2.0 licence — free for any use, including commercial. It's the first Mistral model to fold reasoning (Magistral), image understanding (Pixtral) and agentic coding (Devstral) into one model: a 119-billion-parameter mixture-of-experts design that activates only ~6 billion parameters per token, with a 256k context window and a toggle for reasoning depth.
What it does
One model, three jobs: think through problems, read images and documents, write and run code. Mistral claims 40% lower latency and triple the throughput of Small 3, with benchmark results matching or beating open rivals like GPT-OSS 120B. Available via the Mistral API, Hugging Face, NVIDIA NIM containers, and the frameworks local-AI users already run (vLLM, llama.cpp).
Pricing for Australians
The model is free. What you pay for is somewhere to run it: your own GPU hardware, an Australian cloud GPU instance, or Mistral's API at US$0.15 per million input tokens and US$0.60 output — roughly A$0.21 and A$0.84 at 1.40 (verified September 2026), ex-GST. For the self-hosted route, the MoE design helps — but this is still a serious model needing workstation- or server-class memory — not a laptop toy.
Australian Fit
This is the point. Run Small 4 on hardware in your office or in an Australian data centre and client data never crosses a border — the cleanest Privacy Act story available, particularly for legal, finance and health practices. Apache 2.0 means no lawyer-bait licence terms. No local support, of course; you are the support. 4/5.
Early read
Good: the unification matters — previously you'd juggle three Mistral models; the licence is genuinely free; image + reasoning + code in one onshore-capable package is unique at this quality level. Not so good: the hardware barrier excludes most small businesses from self-hosting; quality still trails the frontier closed models; and "you are the support" cuts both ways when something breaks at quarter-end.
Early verdict
For Australian firms with data that cannot leave the country — and an IT capability or partner to match — Mistral Small 4 became the default open model to evaluate in March. Everyone else should use it the boring way: via API, priced like any other model.
| Criterion | Score |
|---|---|
| Value for Money | 5/5 |
| Ease of Use | 2/5 |
| Feature Depth | 4/5 |
| Australian Fit | 4/5 |
| Reliability & Support | 3/5 |
| Total | 18/25 ⭐⭐⭐⭐ Very Good |
First-look score — provisional until our full hands-on review.
Update history
- 20 Sep 2026 — Staleness check. Re-verified the licence, the architecture and the API price against Mistral's launch post, model list and pricing page. All unchanged: Apache 2.0, 119B total parameters with ~6B active per token (8B with embedding and output layers), 256k context, configurable reasoning effort. The mixture of experts is 128 experts, 4 active per token. Small 4 is still listed among Mistral's latest models, so nothing has superseded it. One new detail we can now state. The API price is US$0.15 per million input tokens and US$0.60 output — about A$0.21 and A$0.84 at 1.40 (verified September 2026), ex-GST. The review previously said only "usage-priced in EUR/USD". The staleness flag has been added at 187 days (house voice rule 8). Score unchanged at 18/25. Launch post · Pricing · Wayback: capture pending (archive.org unavailable 20 Sep 2026)
- 17 Mar 2026 — First reviewed.
Frequently asked questions
What is Mistral Small 4?
A free Apache 2.0 model from French lab Mistral AI, released 16 March 2026. It's the first to fold reasoning, image understanding and agentic coding into one 119-billion-parameter mixture-of-experts model with a 256k context window.
Can Australian businesses run it locally?
Yes. Run it on hardware in your office or an Australian data centre and client data never crosses a border — the cleanest Privacy Act story available, especially for legal, finance and health practices. But you need workstation- or server-class memory; it's no laptop toy.
Is it free?
The model itself is free under Apache 2.0, including for commercial use. What you pay for is somewhere to run it: your own GPU, an Australian cloud GPU instance, or Mistral's API (usage-priced in EUR/USD).
Does it change our recommendation?
For Australian firms with data that cannot leave the country and the IT capability to match, this became the default open model to evaluate in March. Everyone else should use it the boring way — via API, priced like any other model.
Read next
- Category guide: Best Local & Private AI Tools in Australia 2026
- AnythingLLM Review Australia 2026
- GPT4All Review Australia 2026
Sources
- Mistral AI — Introducing Mistral Small 4: https://mistral.ai/news/mistral-small-4/ (primary)
- Hugging Face — Mistral AI models: https://huggingface.co/mistralai (primary distribution)
Price Watch · Free
Know before your AI bill changes
A short email when AI tools change what they charge Australians: the new price in AUD, what it means for you, and whether to switch. No hype, no spam, unsubscribe any time.