Mistral Word Limit by Model (2026)
Mistral is Europe's leading AI lab and the practical choice for teams that need GDPR-compliant, EU-hosted AI. Here's what each Mistral model accepts and where they fit in the 2026 landscape.
Quick Answer
Mistral Large 3 (Dec 2025) accepts 256,000 tokens (~192K words). Mistral Medium 3.5 and Mistral Small 4 also accept 256,000 tokens. Mistral Nemo (open-weight, now deprecated on the Mistral API) accepts 128,000 tokens. Several Mistral models are open-weight under Apache 2.0, making them attractive for EU self-hosting and on-premise deployments that other open-weight options (Llama) don't legally allow in the EU.
Mistral context windows by model
| Model | Input tokens | Input words | License |
|---|---|---|---|
| Mistral Large 3 | 256,000 | ~192,000 | Open weight |
| Mistral Medium 3.5 | 256,000 | ~192,000 | Modified MIT (open weights) |
| Mistral Small 4 | 256,000 | ~192,000 | Apache 2.0 (open) |
| Mistral Large 2 (retired) | 128,000 | ~96,000 | Mistral Research License |
| Mistral Medium 3 (deprecated) | 128,000 | ~96,000 | Proprietary |
| Mistral Small 3 (deprecated on the API) | 32,000 | ~24,000 | Apache 2.0 (open) |
| Mistral Nemo (deprecated on the API) | 128,000 | ~96,000 | Apache 2.0 (open) |
| Codestral | 32,000 | ~24,000 | Mistral NPL |
| Mixtral 8x22B (deprecated on the API) | 64,000 | ~48,000 | Apache 2.0 (open) |
Specs from Mistral AI documentation, September 2026. Some open-weight models have additional extended-context variants available through the community.
Why Mistral matters in Europe
The EU AI Act and data residency rules created a real problem for European companies using American AI. OpenAI, Anthropic, and Google all process data through US infrastructure. Self-hosting Llama 4 is explicitly excluded for EU companies under its Community License. That leaves most enterprises with a narrow set of options.
Mistral is headquartered in Paris, hosts data in EU-based infrastructure, complies with GDPR natively, and offers open-weight models under Apache 2.0 (no commercial restrictions). For EU banks, healthcare providers, and government contractors, Mistral is often the only viable frontier-tier option. Mistral also operates on Microsoft Azure with EU sovereign-cloud configurations for regulated workloads.
Mistral Large 3 vs the competition
On context length, Mistral Large 3's 262K sits mid-pack in 2026: ahead of Claude Haiku's 200K, well behind Gemini's 1M or Llama Scout's 10M. Large 3 is a sparse mixture-of-experts model with 675B total parameters and about 41B active per token, and it is competitive on reasoning, coding, and multilingual tasks — especially strong on European languages (French, German, Spanish, Italian, Dutch).
Where Mistral falls behind: the frontier reasoning tier (GPT-5.6 Sol, Claude Fable 5.1, Gemini 3.1 Pro), aggressive context extension (the 10M Scout tier), and dominant market share in consumer AI tools. For regulated enterprise work, Mistral usually wins anyway because those competitors can't be deployed in the first place.
Open-weight options that actually work
Mistral Nemo and Mistral Small 3 are released under Apache 2.0, which is genuinely permissive — no clauses about monthly active users, no EU exclusions, no commercial restrictions. This is unusual for a lab with a commercial product. Practical hardware requirements:
- Mistral Small 3 (24B params): Runs on a single high-end consumer GPU (RTX 4090 with INT4 quantization). Full 32K context.
- Mistral Nemo (12B params): Runs on a single RTX 3090 or 4090. 128K context window is remarkable for this size class.
- Mixtral 8x22B: Requires 4x A100 or similar for FP16. Strong for enterprise self-hosted workloads.
For teams that need true data isolation (medical records, legal documents, financial data), Mistral Small 3 on a local workstation gives you a competitive LLM that never sends a token anywhere. This is a capability Llama 4 blocks for EU users and Claude/GPT simply don't offer.
Pricing
API pricing as of September 2026 (verify current rates with Mistral):
- Mistral Large 3: ~$0.50 / M input, ~$1.50 / M output
- Retired or deprecated models (Large 2, Medium 3, Small 3, Nemo): no longer listed on Mistral's current model cards, so no current API price is shown here. Mistral Large 3, Medium 3.5 and Small 4 are the replacements.
- Newer tiers (Large 3, Medium 3.5, Small 4): Medium 3.5 is $1.50 / M input and $7.50 / M output, Small 4 is $0.15 / M input and $0.60 / M output
- Self-hosted (all open models): Infrastructure cost only, no per-token charge
Mistral Large 3 at about $0.50 per million input is more expensive than DeepSeek V4.1 Flash ($0.30 at peak) but far cheaper than Claude Opus 5.5 ($4) or GPT-5.6 Sol ($5). For EU-regulated workloads the price premium often pays for itself the first time your compliance auditor asks where the data is processed.
Check your prompt cost across every model
AI Prompt Word CounterFAQ
What is Mistral's word limit?
Mistral Large 3, Medium 3.5 and Small 4 accept about 192,000 words (256K tokens). The retired Large 2 and deprecated Medium 3 accepted about 96,000 words (128K tokens), and Nemo, the open-weight 12B model, supports 96,000 words of input.
Is Mistral GDPR-compliant?
Yes. Mistral is EU-headquartered with EU-hosted infrastructure and native GDPR compliance. This is the key reason many European enterprises choose Mistral over US-based alternatives.
Can I run Mistral models locally?
Yes for the open-weight models. Mistral Small 3 runs on a high-end consumer GPU. Nemo runs on a single RTX 3090 or 4090. Mixtral 8x22B needs datacenter hardware. Large 2 and Medium 3 are retired or deprecated on the API.
How does Mistral Large 2 compare to GPT-4?
Mistral Large 2 was retired on March 30, 2025 (Large 3 replaces it). It was roughly competitive on benchmarks. Slightly behind on English reasoning, slightly ahead on European languages. Context window (128K) is smaller than GPT-4.1's 1M but sufficient for most workloads.
Is Mistral good for coding?
Yes. Codestral is Mistral's dedicated coding model with 128K context. Codestral-Mamba uses state-space architecture for longer effective context on code tasks. Competitive with GPT-4 class for most coding work.