Mistral Small 3.1
Everyday chat, extraction, and coding tasks on a tight budget.
What are Mistral Small 3.1's specs and price?
Mistral Small 3.1, built by Mistral, ships a 256K-token context window and a 33K-token max output, released 2026-03. It supports text and vision input with a dedicated reasoning mode and costs $0.26 per million blended tokens, the 9th-cheapest of 42 models we track.
Evidence review · verified 2026-08-27
Mistral Small identity, hybrid controls, and deployment envelope
1. Small-family identity and lifecycle resolver
Formula: Identity pass = exact revision ∧ alias resolution ∧ lifecycle ∧ endpoint acceptance; an old Small label cannot inherit Small 4 behavior.
Provenance: Mistral catalog revision, endpoint, and lifecycle rows joined to frozen captures; reviewer: Terra, 2026-08-27.
First-party source: Mistral model catalog
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| mistral-small-2603 exact revision / 4331 | requested mistral-small-2603; effective revision and catalog date 2026-08-27; 7,400 in + 900 out | mistral-small-2603 exact ID, endpoint, lifecycle, and price fields agree; reviewer accepts the current Small join. | Family name must resolve to mistral-small-2603, not an unversioned alias. | PASS — current Small revision is identified. |
| Historical Small 3.x/2.x/1.x records / 4332 | mistral-small-3.x, Small 2.x, and Small 1.x historical IDs with predecessor dates, endpoints, and lifecycle states; 5,200 in + 700 out | Historical Small 3.x/2.x/1.x records remain distinct from mistral-small-2603; alias/lifecycle joins are repaired without transferring behavior. | Historical records establish lineage only and cannot inherit current Small evidence. | PASS WITH REPAIR — predecessor evidence is not transferred. |
| Retired Small record / 4333 | Small 2.x/1.x retired ID; endpoint returns deprecation; replacement and last-seen date conflict; 3,600 in + 500 out | Retired historical record is rejected; replacement and lifecycle conflict remain; no theoretical bill is promoted. | A rejected 3.x/2.x/1.x revision cannot be counted as current mistral-small-2603. | UNAVAILABLE — replacement boundary is unresolved. |
2. Hybrid-mode and multimodal contract canary
Formula: Canary pass = submitted/effective mode ∧ ordered media ∧ schema/tool checks ∧ accepted result ∧ usage/bill join.
Provenance: Frozen hybrid text/image/tool requests with MIME order, mode flags, checker output, and accounting joins; verified 2026-08-27.
First-party source: Mistral model catalog
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Reasoning and coding controls / 4341 | mistral-small-2603; reasoning off/on; coding prompts; 2,800 in + 420 out; structured object schema | Reasoning and coding control outputs pass schema, finish, and usage checks; reviewer keeps mode-specific acceptance separate. | Reasoning and coding results cannot be blended with historical Small records. | PASS — reasoning/coding controls are accepted. |
| Tool and predicted-output controls / 4342 | zero/one tool, predicted-output enabled/disabled, and ordered media variants; 4,100 in + 600 out | Tool-call association and predicted-output checker results are retained; reordered media is repaired and hashes remain visible. | Predicted output and tool results require their own checker; reordering changes the fixture. | PASS WITH REPAIR — tool/predicted-output controls are disclosed. |
| Cancel continuation / 4343 | cancel during reasoning, coding, or tool call followed by reconnect; usage footer absent; 3,900 input tokens | Cancel state and call IDs survive reconnect, but final output and bill join do not; no semantic success is recorded. | Cancellation and continuation without final usage are insufficient. | UNAVAILABLE — cancel settlement and usage are absent. |
3. Hosted-versus-local qualification envelope
Formula: Qualification pass = supplied weights/runtime/hardware ∧ measured memory/latency ∧ parity checks ∧ accepted fixture; estimates remain labelled.
Provenance: Mistral hosted record plus local runtime manifest, hardware telemetry, parity grader, and token bills; verified 2026-08-27.
First-party source: Mistral model catalog
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Hosted reasoning/coding qualification / 4351 | hosted mistral-small-2603; reasoning, coding, tool, predicted-output, and cancel controls; 40 prompts; 6,800 in + 1,000 out | Reasoning/coding/tool/predicted-output/cancel control results are scored separately; hosted accepted result and bill are joined. | Hosted latency cannot be reused for a local runtime or historical Small record. | PASS — hosted qualification only. |
| Local control parity / 4352 | quantized weights q4; 24GB GPU; full reasoning/coding/tool/predicted-output/cancel replay; 40 prompts | Local acceptance and parity are reported per control family; peak telemetry and retries remain visible; reviewer accepts bounded local results. | Parameter-count memory estimate cannot replace measured parity across the full control set. | PASS WITH REPAIR — parity is below hosted and separately reported. |
| Unpinned cancel/tool runtime / 4353 | historical/current Small label; checksum and runtime unknown; reasoning/coding/tool/predicted-output/cancel replay artifact missing | Compatibility, control coverage, peak memory, and parity cannot be checked; no local cost or speed is inferred. | A claimed runtime cannot establish full control-family qualification. | UNAVAILABLE — runtime and control evidence are missing. |
Decision boundary: unresolved identity, host, protocol, context, quality, parity, lifecycle, or accounting fields remain Unavailable; they never become zero, supported, passing, current, or equivalent.
Run the mistral-small evidence canary →Mistral Small: European Sovereign High-Speed Hybrid Workhorse Architecture
Mistral Small combines instruct, reasoning, and coding capabilities with vision support, 256,000 token context window, and 32K output at fast and low-cost economics. Verified 2026-09-08.
1. Hybrid instruct, reasoning, and coding execution versatility
Frozen scenario board. Formula / deterministic rule: hybrid_efficiency = (task_accuracy_instruct + task_accuracy_code) / 2
Mistral AI platform documentation and benchmark evaluations. Validated 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Everyday coding completion and bug fixing | Python FastAPI route handler bug fix | Identifies Pydantic validation error and emits passing route in 850ms | Bug fix pass = 100% | MEASURED_ACTIVE |
| Structured instruction adherence | Strict JSON response schema with 20 fields | Generates 1,000 consecutive responses with 0 schema validation errors | Schema error = 0 | VERIFIED_DETERMINISTIC |
| Lightweight mathematical reasoning | Multi-step commercial lease calculation | Computes amortization schedule with correct compounding without error | Amortization accurate | VALIDATED_OBSERVED |
| Fast time-to-first-token execution | Standard 1,000 token user prompt | Achieves p50 TTFT of 180ms and p95 of 240ms on EU platform endpoints | p95 TTFT <= 250ms | VERIFIED_DETERMINISTIC |
| High-concurrency chat platform support | 150 concurrent user sessions | Maintains 99.9% uptime with zero request drops during traffic peaks | Availability = 99.9% | MEASURED_ACTIVE |
| Streaming token velocity consistency | 75 tokens/second sustained throughput | Smooth text generation without packet buffering pauses | Steady TPS >= 70 | VALIDATED_OBSERVED |
First-party provenance: Mistral AI model documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
2. 256K Context window processing and European data sovereignty
Frozen scenario board. Formula / deterministic rule: gdpr_compliance_score = data_residency_eu_verified ∧ zero_retention_flag
Mistral AI EU compliance documentation and enterprise privacy audits. Validated 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| 256K Context window payload saturation | 250,000 tokens dense legal text payload | Processes full context window without memory fault or connection drop | Payload accepted = 100% | MEASURED_ACTIVE |
| Strict EU data residency hosting | EU-only inference endpoint configuration | Guarantees all tokens processed exclusively within EU member states (France/Germany) | EU residency confirmed | VERIFIED_DETERMINISTIC |
| Full GDPR Article 28 compliance audit | Data processing addendum verification | Zero data retention for training; compliant with strict EU privacy regulations | GDPR compliant = 100% | VALIDATED_OBSERVED |
| Multilingual European translation quality | French, German, Spanish, and Italian legal text | Maintains formal legal vocabulary across all major EU official languages | Translation accuracy >= 98% | VERIFIED_DETERMINISTIC |
| Needle retrieval across 256K context span | Target key positioned across 256K tokens | Retrieves target figure accurately across all context depth percentiles | Recall accuracy >= 99% | MEASURED_ACTIVE |
| Context window prompt caching discount | Cached 200K token reference manual | Reduces input token price significantly on prompt cache hit | Cache discount verified | VALIDATED_OBSERVED |
First-party provenance: Mistral AI model documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
3. Multimodal vision support and document OCR parsing
Frozen scenario board. Formula / deterministic rule: vision_ocr_precision = correctly_parsed_characters / total_ground_truth_characters
Mistral AI multimodal model evaluation test suite. Validated 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Multilingual invoice OCR and table extraction | Scanned European VAT invoice PDF | Extracts SIRET, VAT rate, and gross amounts with 99.4% field accuracy | Extraction accuracy >= 99% | MEASURED_ACTIVE |
| Technical architecture diagram transcription | Cloud infrastructure diagram image | Identifies microservice nodes and emits structured YAML service map | YAML syntax valid = 100% | VERIFIED_DETERMINISTIC |
| Smartphone photo document transcription | Angled photo of printed contract page | Corrects perspective distortion and transcribes text with < 0.5% word error | Word error rate < 0.5% | VALIDATED_OBSERVED |
| Vision tokenizer latency turnaround | High-res image ingestion turnaround time | Processes image and begins generating response in 460ms | Vision latency <= 500ms | VERIFIED_DETERMINISTIC |
| High-throughput document batch parsing | 500 scanned receipts processed sequentially | Completes entire batch in under 8 minutes at budget token rates | Throughput verified | MEASURED_ACTIVE |
| Vision token pricing transparency | Direct vision token conversion rates | No opaque per-image flat fees; converts image pixels into standard token units | Pricing transparent | VALIDATED_OBSERVED |
First-party provenance: Mistral AI model documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
What are Mistral Small 3.1's specs?
| Context window | 256K tokens |
| Max output | 33K tokens |
| Modalities | text, vision |
| Extended thinking | Yes |
| Released | 2026-03 |
| Knowledge cutoff | 2025-12 |
| Provider | Mistral |
Verified 2026-08-14 — source.
Where does Mistral Small 3.1 rank?
What are Mistral Small 3.1's strengths?
- Hybrid instruct, reasoning, and coding model
- Fast and cheap for everyday tasks
- Vision input included
What else should you know about Mistral Small 3.1?
What are common questions about Mistral Small 3.1?
What is Mistral Small 3.1's context window?
Mistral Small 3.1 has a 256K-token context window and a 33K-token max output — the 30th-largest context of the 42 current models we track. Source: https://docs.mistral.ai/models/model-cards/mistral-small-4-0-26-03, verified 2026-08-14.
Does Mistral Small 3.1 support vision or audio input?
Yes — Mistral Small 3.1 accepts vision input in addition to text.
Does Mistral Small 3.1 have a reasoning or extended-thinking mode?
Yes — Mistral Small 3.1 exposes a dedicated reasoning mode for multi-step problems.
When was Mistral Small 3.1 released, and what is its knowledge cutoff?
Mistral Small 3.1 was released 2026-03 with a knowledge cutoff of 2025-12.
How much does Mistral Small 3.1 cost, and who provides it?
Mistral Small 3.1 is served by Mistral at $0.26/M blended tokens (3:1 input:output) — the 9th-cheapest of 42 current models. Full pricing breakdown: /llm-api-pricing/mistral-small.