证据: 基准测试结果由厂商自行报告
Mistral Large 4 1T-A52B — 显存、速度与本地部署
作者： Jakub Rusinowski · 最后更新：
总参数约 1T，每个 token 激活 52B（不含嵌入层和输出层为 49B）：128 个路由专家，激活 4 个，另有 1 个共享专家；采用带稀疏注意力索引器的潜在注意力，100 万 token 上下文，1.6B 视觉编码器。推理强度可设为 low、medium 或 high，无法关闭。Mistral 表示该模型在其位于欧洲的自有数据中心、使用 3,800 块 NVIDIA Grace Blackwell GPU 从零训练。官方权重为 FP8 和 NVFP4；llama.cpp 维护者发布的 GGUF 从 298 GB（IQ2_XXS）到 2,099 GB（BF16）。
Mistral Large 4 1T-A52B 在 Q4_K_M 下约需 636 GB 显存——量化权重加框架开销，不含 KV 缓存。在 Apple Silicon 上，这部分来自统一内存。
运行 Mistral Large 4 1T-A52B
Mistral Large 4 1T-A52B is not practical on a single consumer GPU: it needs about 635.6 GB at Q4_K_M, more than any single device in our catalogue holds with headroom. How multi-GPU inference works, or rent a multi-GPU server by the hour:
如何运行 Mistral Large 4 1T-A52B
Ollama lists Mistral Large 4 only as the cloud tag mistral-large-4:cloud, which runs on Ollama's servers, not your machine. Locally, use llama.cpp (mainline since 29 October 2026) with the ggml-org GGUFs, or vLLM / SGLang on a multi-GPU server.
规格
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- 参数量
- 1.05 Trillion (52B active) + 1.6B vision encoder
- 上下文窗口
- 1,048,576
- 架构
- Mixture-of-Experts (128 routed + 1 shared, 4 active), MLA with sparse-attention indexer
- 提供商
- Mistral AI
- 许可证
- Mistral Ecosystem Open-Weight License v26.10
- Availability
- Weights downloadable
- Practical deployment
- Private cloud / datacenter cluster
- 规格量化
- Q4_K_M
- Vision projector
- +3.4 GB when you send images (not included above)
- 系统内存
- 1024 GB
- 记录更新于
- 2026-11-11
Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.
限制与注意事项
- Datacentre-only in practice: every expert must be resident. The one single-machine route is a 2-bit GGUF (IQ2_XXS 297.7 GB or Q2_K 380.7 GB), which the llama.cpp maintainers say fits 512 GB of unified memory; on Apple Silicon that is close to the default GPU memory limit.
- No KV-cache or speed estimate here: the cache is shared across layers and a sparse-attention index adds its own, which our model does not yet express. Hugging Face measured about 15 tokens/s generation with Q2_K on one Mac Studio M3 Ultra (512 GB).
- On Mistral's API the model is still labelled Public Preview (docs version 26.10).
- Every benchmark in the launch material is vendor-reported; preview and final numbers differ and must not be mixed.
质量与使用场景
评分由模型作者或独立评测方发布——衡量质量而非吞吐量，并非我们实测。
Mistral Large 4 1T-A52B — 常见问题
How much VRAM does Mistral Large 4 1T-A52B need?
About 636 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. We publish no context-length figure or speed for it: its attention shares the KV cache across layers and adds a sparse-attention index cache, which our memory and speed model cannot yet express, so we state the published weights only. Apple Silicon counts unified memory toward the same figure.
Does Mistral Large 4 1T-A52B run on an RTX 4090 (24 GB)?
Its weights and overhead come to about 636 GB at Q4_K_M, but we do not publish a fit verdict or a speed for it: its attention shares the KV cache across layers and adds a sparse-attention index cache, which our memory and speed model cannot yet express, so we state the published weights only.
How do I run Mistral Large 4 1T-A52B locally?
Ollama lists Mistral Large 4 only as the cloud tag mistral-large-4:cloud, which runs on Ollama's servers, not your machine. Locally, use llama.cpp (mainline since 29 October 2026) with the ggml-org GGUFs, or vLLM / SGLang on a multi-GPU server. Running the published tag would send your prompts to a hosted GPU rather than your own machine.