by Mistral AI · Mistral Large family · best for EU-sovereign open-weight frontier generalist
Mistral Large 3 (release 25.12, shipped 2 December 2025) is the French lab's open-weight frontier flagship and the centerpiece of its EU-sovereignty pitch. It is a 675B-parameter granular Mixture-of-Experts with only 41B active per token (39B language model plus a 2.5B vision encoder), trained from scratch on 3,000 NVIDIA H200s, released under a genuine Apache 2.0 license. The single sentence a buyer needs: it is the strongest permissively-licensed, self-hostable generalist model a regulated European enterprise can run inside its own VPC, at roughly a quarter of a US flagship's API price.
| Benchmark | Score | Source |
|---|---|---|
| MMLU | 85.5% | intuitionlabs.ai 2025-12-02T00:00:00.000Z |
| Artificial Analysis Index | 23 | artificialanalysis.ai 2026-05-28T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“The only frontier-adjacent model I can put on-prem under Apache 2.0 and tell my DPO it never leaves the EU — sovereignty as a product feature.”
For a one-to-three-year platform bet in Europe, Large 3 is the strategic default. Apache 2.0 weights plus EU-default residency plus SOC 2/ISO 27001 remove the two biggest enterprise blockers — vendor lock-in and Cloud Act exposure — in one model. Multi-cloud managed availability (Bedrock, Azure, watsonx) makes failover real. The trade-off is capability: this is a strong generalist, not a US-frontier reasoner, so analytical and hard-coding workloads still benchmark better elsewhere. For any architecture decision where EU deployment is on the table, it is the obvious open option; for pure capability with no residency constraint, it is not.
“Mistral owns the 'European AI champion' narrative, and Large 3 is the flagship that makes the sovereignty positioning credible rather than rhetorical.”
Strategically, Large 3 is positioned precisely where US and Chinese labs are weak: a Western, GDPR-native, Apache-licensed frontier model. That is a defensible moat in EU public sector, defence, and regulated finance, segments where neither OpenAI nor DeepSeek can compete on residency. Against Llama 4 it wins decisively on European-language quality and license clarity; against US closed models it wins on sovereignty and price. The risk is that "sovereignty" is a moat in Europe and a niche elsewhere — in US-only procurement the value proposition mostly evaporates. Market timing is excellent: EU AI Act obligations land August 2026, making the compliance story sharper.
“$0.50 in, $1.50 out for an open-weight frontier model — and if volume justifies it, I amortise GPUs instead of renting margin from a US vendor forever.”
The unit economics are excellent. At $0.50/$1.50 Large 3 is 3-5x cheaper than US flagships at the generalist tier, with a $0.05 cached-input rate and ~50% batch discount layered on. The deeper financial argument is self-host: Apache 2.0 converts variable per-token spend into fixed GPU capex, which makes annual budgeting predictable for sustained high-volume workloads. The break-even is real only at scale — a single 8x-H100 node is a meaningful capital outlay — so for low volume the API is cheaper. Cache and batch discipline matter given the heavy-node footprint. Strong value, best realised at scale.
“OpenAI-compatible JSON, predictable tool calls, and an actual self-host escape hatch — I can prototype on La Plateforme and ship on vLLM without a rewrite.”
The API is clean and OpenAI-compatible; JSON mode and function calls behave predictably, and SDK quality has caught up to the leaders. The killer feature for a builder is portability: prototype against the hosted endpoint, then self-host the identical Apache 2.0 weights on vLLM or TGI with no code rewrite. Multilingual code-comment and identifier handling beat Llama 4 for European-language repos, and 256K context is genuine. Downsides: La Plateforme's docs are thinner than Anthropic's, function-calling edge cases occasionally surprise, and observability tooling lags OpenAI's. A strong default for any EU-deployed agent project.
“In French and German it's the best open model I've used; in English it's a competent half-step behind Claude and GPT.”
In Le Chat the model is fast enough, accurate in European languages, and more willing to engage nuanced topics than the most heavily safety-tuned US models — refusals are moderate and consistent. For a French or German user this is the best conversation partner among open models. English conversational quality sits a tier below Claude or GPT-5: a touch more literal, less warm. Vision answers work but feel less polished than the leaders. Latency at ~51 tps is fine for chat but not snappy. Overall a strong daily driver for European-language users, merely good for English-first ones.
“Mistral shipped comparison charts and an Apache license, then quietly withheld the standard benchmarks — AA Index 23 against a 36 median tells the part the launch deck didn't.”
The sovereignty and license story is genuine and the price is real. The capability story is where I push back. Mistral published MMLU and LMArena bragging rights but withheld GPQA Diamond, MMLU-Pro, AIME, LiveCodeBench, and SWE-bench at launch — the benchmarks where it would be measured against US frontier models. Independent aggregation fills the gap: Artificial Analysis puts the Intelligence Index at 23 versus a ~36 frontier median, i.e. clearly below GPT-5.x and Claude/Gemini frontier. The "frontier" label is marketing; "best open generalist with EU residency" is the honest claim. Buy it for sovereignty and price, not because it out-thinks the US leaders.
The full research notes behind this review — verified against primary sources.
Large 3 is a granular sparse Mixture-of-Experts. Mistral's own card states 675B total parameters split as a 673B MoE language model with ~39B active plus a 2.5B vision encoder, for ~41B active per forward pass. The number of experts is not disclosed. It was trained from scratch on 3,000 H200 GPUs. The model ships natively in FP8, with NVFP4 and BF16 variants published; FP8 fits a single B200/H200 node while NVFP4 runs on H100/A100 nodes. Tokenization uses mistral_common (>= 1.8.6). Attention mechanism, layer count, training-token count, and vocab size are undisclosed — typical for Mistral, which discloses parameter counts for open-weight models but withholds deeper internals.
Large 3 is a generalist multimodal MoE tuned for instruction following, code generation, multilingual dialogue, document understanding, and tool use. The granular routing activates ~6% of parameters per token, so it draws on the breadth of a 675B pool at the cost of a mid-size dense model (cap_coding 7.5, cap_agentic 7.5). Its standout axis is multilingual quality (cap_multilingual 9.0): French, German, Spanish, Italian, Portuguese, and Dutch output reads natively, with competent Chinese, Japanese, Korean, and Arabic across 40+ languages. The 256K context is real and usable (cap_long_context 8.0). Native vision (cap_vision 7.0) handles charts, screenshots, and documents but trails Gemini and GPT-class vision on hard document benchmarks. It has no default chain-of-thought, so it trails dedicated reasoning models on multi-step math and PhD-science (cap_reasoning 7.0, cap_math 6.5). Tool use, JSON-mode structured output, and function calling are reliable (cap_function_calling 8.0). It has no native real-time retrieval (cap_realtime_data 0.0).
| Benchmark | Score | vs Predecessor | vs Top Competitor | Source |
|---|---|---|---|---|
| MMLU | 85.5% | +~1.5pp vs Large 2.1 | trails GPT-5.x (~89%) | IntuitionLabs |
| Artificial Analysis Index | 23 | new methodology | below frontier median (~36) | Artificial Analysis |
| LMArena rank | #2 OSS non-reasoning | up | behind larger OSS models on reasoning | Mistral |
Mistral published comparison charts at launch but withheld most numeric standard-benchmark scores (MMLU-Pro, GPQA Diamond, AIME, LiveCodeBench, SWE-bench). Benchmark coverage is partial — only the values with a verifiable source are recorded; the rest are null.
Artificial Analysis measures Large 3 at 50.7 output tokens/sec with a 1.09s time-to-first-token — squarely in the medium latency tier and notably slower than the throughput-optimized smaller Mistrals. For a 675B MoE this is expected; interactive chat feels responsive but not snappy, and the model is best deployed for backend pipelines, batch document processing, and agent loops rather than sub-second UX.
| Surface | Cost | Notes |
|---|---|---|
| API input | $0.50 / 1M tok | La Plateforme |
| API output | $1.50 / 1M tok | La Plateforme |
| Cached input | $0.05 / 1M tok | cache read |
| Batch (in/out) | $0.25 / $0.75 | ~50% async discount |
| Direct UI | EUR 14.99/mo (~$15) | Le Chat Pro |
| Free tier | ~25 msg/day (Le Chat); daily token quota (La Plateforme) | no card required |
| Self-host | Apache 2.0 | weights on Hugging Face |
| Cloud | Bedrock, Azure AI Foundry, IBM watsonx | managed |
This is the strongest deployment story in the Mistral lineup. Apache 2.0 weights are published on Hugging Face (FP8/NVFP4/BF16); self-hosting needs a single high-memory node (~320GB+ VRAM, i.e. 8x H100/H200 or B200, less at NVFP4). Managed availability spans Amazon Bedrock, Azure AI Foundry, and IBM watsonx. La Plateforme is EU-hosted by default — the data-residency wedge — with US options. For a regulated French, German, or Dutch enterprise, Large 3 is the canonical choice: frontier-adjacent capability that can run on-prem under a permissive license, satisfying data-residency rules and avoiding US Cloud Act exposure.
Mistral is a French company natively subject to GDPR, holds SOC 2 Type II and ISO 27001/27701, and positions itself as aligned with the EU AI Act. API inputs are retained 30 rolling days for abuse monitoring and are not used for training unless the customer opts in; Zero Data Retention is available and removes the 30-day window. EU data residency by default is the central enterprise selling point. Content moderation is not built into the model but is offered as a separate Mistral Moderation API (mistral-moderation-2603, 9 categories plus jailbreak detection). Refusal calibration is moderate and consistent — less aggressively safety-tuned than US frontier labs.
SDKs in Python and TypeScript/JavaScript; integrations with LangChain, LlamaIndex, Vercel AI SDK, and Haystack. Powers Le Chat and Mistral AI Studio. Open weights drive a growing self-host and fine-tune community on Hugging Face (FP8/NVFP4/BF16/GGUF derivatives). Popularity is growing rather than dominant — strong in Europe, narrower elsewhere.
Yes — genuine Apache 2.0 (verified on the Hugging Face card), permitting commercial use, modification, and redistribution with no revenue threshold. This is unlike Medium 3.5 and Devstral 2, which carry a modified-MIT revenue restriction.
FP8 needs a single high-memory node (~8x H100/H200, ~320GB+ VRAM); NVFP4 lowers the bar to H100/A100 nodes. Break-even versus the API arrives only at sustained high volume.
La Plateforme is EU-hosted by default with US options; 30-day abuse-monitoring retention, no training on inputs unless you opt in, and Zero Data Retention available.
On generalist tasks it is competitive; on hard reasoning (GPQA, AIME) and English polish it trails GPT-5.x and Claude. AA Index 23 vs ~36 frontier median.
Not by default — there is no chain-of-thought toggle on Large 3. A reasoning variant was announced as coming; for extended thinking today, route to Magistral or Medium 3.5.
Amazon Bedrock, Azure AI Foundry, and IBM watsonx, plus self-host and La Plateforme.
SOC 2 Type II, ISO 27001/27701, GDPR-native, EU AI Act aligned (non-high-risk obligations land August 2026).
Comparable open-weight MoE tier; weaker on European languages and license clarity, stronger US ecosystem and tooling support.
Cheaper and comparable on English coding; weaker EU-multilingual quality and a less clean permissive-license story; Chinese-origin residency is a non-starter for EU-sovereignty buyers.
Closed weights, 2-5x the price, clearly stronger on reasoning and English polish, but no EU data-sovereignty or self-host story — the inverse trade-off to Large 3.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-05-27