Mistral Large 3
Mistral Large 3 is a large language model developed by French AI company Mistral AI, belonging to its flagship model series. The model provides strong multilingual processing capabilities, supporting complex reasoning, code generation, and long-text comprehension, and is suited for enterprise-level natural language processing tasks. Its design emphasizes a balance between high performance and efficiency, and it is available via API and on-premises deployment for developers and commercial applications.
Core Facts
Timeline (last 90 days)
When the context approaches 512K tokens, the prefill-decode disaggregated architecture ensures that users still receive the first response token within 800 milliseconds after reading an entire specification.
For large enterprises processing more than 500 million tokens per day, using an open-weight Mistral Large 3 local deployment can reduce the effective cost per million tokens to below $0.35, saving more than 80% compared with closed-source competitor APIs.
The recommended distributed parallelism strategy for Mistral Large 3 uses expert parallelism (EP=8 or 16) and tensor parallelism (TP=4 or 8), and requires an InfiniBand / RoCE v2 + RDMA network with at least 400 Gbps per port.
Mistral Large 3 uses a MoE architecture composed of 128 fine-grained experts, dynamically activating 8 of them per token.
Mistral Large 3 has a total parameter count of 420 billion, and the decoding phase only needs to bear the compute cost and latency of 68 billion active parameters.
Mistral Large 3 supports a native context window of 512,000 tokens, combining block sparse attention with hardware-calibrated rotary position embeddings (RoPE).
In the Needle In A Haystack test, Mistral Large 3 achieved 99.8% retrieval accuracy at any depth quadrant of the half-million-token window.
Mistral Large 3 weights are openly downloadable in FP8/BF16 format and can be deployed locally in a sovereign cloud or a fully offline air-gapped environment.
On the SWE-bench Verified benchmark, Mistral Large 3 achieved a score of 53.8%.
On the Terminal-Bench 4.0 benchmark, Mistral Large 3 achieved a score of 65.2%, ranking second worldwide.
3 more timeline events
All Facts (13)
Mistral Large 4 scored only 38 points on the Artificial Analysis composite intelligence index, ranking sixth among open-source models after excluding closed-source models.
70%UnverifiedMistral Large 3 uses a MoE architecture composed of 128 fine-grained experts, dynamically activating 8 of them per token.
50%UnverifiedMistral Large 3 has a total parameter count of 420 billion, and the decoding phase only needs to bear the compute cost and latency of 68 billion active parameters.
50%UnverifiedMistral Large 3 supports a native context window of 512,000 tokens, combining block sparse attention with hardware-calibrated rotary position embeddings (RoPE).
50%UnverifiedIn the Needle In A Haystack test, Mistral Large 3 achieved 99.8% retrieval accuracy at any depth quadrant of the half-million-token window.
50%UnverifiedMistral Large 3 weights are openly downloadable in FP8/BF16 format and can be deployed locally in a sovereign cloud or a fully offline air-gapped environment.
50%UnverifiedOn the Terminal-Bench 4.0 benchmark, Mistral Large 3 achieved a score of 65.2%, ranking second worldwide.
50%UnverifiedOn an 8x H200/B200 node, Mistral Large 3 achieves a sustained generation speed of 142 tokens/second.
50%UnverifiedThe recommended distributed parallelism strategy for Mistral Large 3 uses expert parallelism (EP=8 or 16) and tensor parallelism (TP=4 or 8), and requires an InfiniBand / RoCE v2 + RDMA network with at least 400 Gbps per port.
50%UnverifiedOn the SWE-bench Verified benchmark, Mistral Large 3 achieved a score of 53.8%.
50%UnverifiedFrom Large 3 to Large 4, total parameters grew by 56% while active parameters grew by only 20%.
50%UnverifiedWhen the context approaches 512K tokens, the prefill-decode disaggregated architecture ensures that users still receive the first response token within 800 milliseconds after reading an entire specification.
50%UnverifiedFor large enterprises processing more than 500 million tokens per day, using an open-weight Mistral Large 3 local deployment can reduce the effective cost per million tokens to below $0.35, saving more than 80% compared with closed-source competitor APIs.
50%