Mistral Large 2 (123B)

Positioning

Mistral Large 2 is Mistral AI's 123B-parameter dense model — the top multilingual model in Mistral's lineup and arguably the strongest open-weight model for European language tasks. Sitting below the massive Mistral Medium 3.5 (675B MoE) in the product stack, Large 2 is the pragmatic flagship: frontier-quality multilingual performance at hardware requirements that fit a single A100 80GB or Mac Studio M3 Ultra. For operators whose workloads span multiple European languages, Mistral Large 2 is the reference pick.

Strengths

• Best-in-class European multilingual performance. Mistral's training pipeline consistently produces the strongest French, German, Spanish, Italian, and Portuguese performance among open-weight models. Large 2 is the reference model for European-language NLP tasks.
• 123B dense architecture: consistent quality across tokens. Every token sees all 123B parameters — no MoE routing variance, no expert collision artifacts. Quality is predictable and consistent.
• Reasonable hardware requirements for a 123B model. Q4 at ~70 GB fits a single A100 80GB or Mac Studio M3 Ultra. Q3 at ~53 GB fits dual RTX 4090 setups. Much more accessible than the 405B+ frontier class.
• Mistral's engineering discipline. Clean architecture, well-tuned tokenizer, predictable serving behavior. Mistral models are known for being well-behaved in production.
• Strong coding and reasoning benchmarks. Competitive with Llama 3.3 70B and Qwen 2.5 72B Instruct on HumanEval, MBPP, and general reasoning tasks. The 123B scale provides meaningful capability headroom over 70B-class models.
• 128K context window. Long-context handling at 123B scale — competitive with the best dense models.

Limitations

• Mistral's commercial license. Not Apache 2.0. Verify the Mistral license terms for commercial deployment, especially for large-scale or redistribution use cases.
• Dense architecture = higher per-token compute than same-scale MoEs. Every token activates all 123B parameters, vs ~22-36B for comparable MoEs. Tok/s is slower than MoE models at similar total parameter counts, though quality is more consistent.
• English performance trails frontier MoEs. While competitive, Mistral Large 2's English benchmark scores are slightly behind DeepSeek V3 and Qwen 3 235B-A22B on English-specific tasks. The model's strength is multilingual breadth, not English-only depth.
• No vision or multimodal capability. Text-only. For vision-language tasks at similar scale, consider Llama 3.2 90B Vision or InternVL 2.5 78B.
• Single consumer GPU path is tight. At Q3 (~53 GB), a single RTX 5090 (32 GB) can't fit this. Dual consumer GPU or workstation-class hardware is required. This is not a single-card model for most operators.

Real-world performance

• vs Mistral Medium 3.5: Large 2 is the pragmatic sibling — similar multilingual DNA at dramatically lower hardware cost (70 GB Q4 vs 380 GB Q4). Medium 3.5 has higher capability ceiling but Large 2 fits on hardware most operators can actually access.
• vs Llama 3.3 70B: Large 2 has 53B more parameters and meaningfully stronger multilingual performance. Llama 3.3 70B wins on English-only depth and ecosystem breadth. Pick by dominant language.
• vs Qwen 2.5 72B Instruct: Large 2 is larger (123B vs 72B) and stronger on European languages. Qwen 2.5 72B Instruct uses the Qwen License (permissive commercial terms) and stronger on Chinese. Pick by language + license preference.

Should you run this locally?

Yes if your workload spans multiple European languages (French, German, Spanish, Italian, Portuguese), you have workstation-class hardware (Mac Studio M3 Ultra at Q4, or dual A100 80GB), and Mistral's commercial license works for your deployment. Mistral Large 2 is the reference model for European multilingual NLP.

No if your workload is English-only (pick Llama 3.3 70B), you need Apache 2.0 licensing (pick Mixtral 8x22B), or you need single-consumer-GPU deployment (pick smaller 70B-class models).

How it compares

• vs Mistral Medium 3.5: Smaller, more deployable sibling. Same multilingual DNA, lower hardware cost.
• vs Llama 3.3 70B: Larger scale (123B vs 70B); stronger multilingual; Llama wins English + ecosystem.
• vs Qwen 2.5 72B Instruct: Large 2 is larger and stronger on European languages; Qwen uses the Qwen License (permissive commercial terms) and stronger on Chinese.
• vs Command R Plus: Both strong multilingual models; Mistral has broader language coverage; Command R Plus has Cohere's RAG pipeline strengths.

Run this yourself

• Datacenter: Single A100 80GB at FP8 with vLLM. Production target: 40-60 tok/s.
• Workstation: Mac Studio M3 Ultra (192 GB) at Q4 — comfortable fit at 12-18 tok/s.
• Dual consumer GPU: 2× RTX 4090 at Q3 with ExLlamaV2 tensor parallelism — functional at 15-25 tok/s.
• Vendor: huggingface.co/mistralai/Mistral-Large-2