{"slug":"deepinfer","name":"DeepInfer","nameZh":"DeepInfer（DeepSeek 官方开源推理框架）","category":"local-inference","categoryName":"Local inference & serving","audience":["ai-era"],"primary":"ai-era","openClosed":"open","pricing":"open-source","license":"Apache-2.0","maintained":"yes","officialUrl":"https://github.com/deepseek-ai/DeepInfer","repo":"https://github.com/deepseek-ai/DeepInfer","latestVersion":null,"latestRelease":null,"stars":18108,"starsAsOf":"2026-10-22","description":"The open-source release of the inference framework behind DeepSeek’s production services (repo created 2026-10-16, ~18k stars in a week): currently DeepSeek-V4.1-Flash only, offering offline inference with the core features for serving at scale — FlashMLA / DeepGEMM / DeepEP kernels, attention DP + MoE EP/EPLB, draft-token decoding (DSpark), async scheduling, batch-invariant and deterministic sampling, structured outputs. Runs on NVIDIA SM100 (Blackwell) and Huawei Ascend 950 only; the KV-cache system, prefill/decode disaggregation and load balancing needed for a full online service are not included","descriptionZh":"DeepSeek 线上大规模服务所用推理框架的开源版（2026-10-16 建仓，一周 1.8 万星）：目前只支持 DeepSeek-V4.1-Flash，提供离线推理和规模化部署的核心能力——FlashMLA / DeepGEMM / DeepEP 等自研算子、Attention DP + MoE EP/EPLB 并行、草稿 token 加速解码（DSpark）、异步调度、批次不变与确定性采样、结构化输出。硬件限 NVIDIA SM100（Blackwell）和华为昇腾 950；完整在线服务所需的 KV 缓存系统、预填充/解码分离和负载均衡不在本次开源范围内","callMethods":[],"platforms":[],"addedAt":"2026-10-22","updatedAt":"2026-10-22","topics":[],"language":"Python","lastPush":"2026-10-21","formerly":null,"packages":[],"recentVersions":[],"alternatives":["ollama","llama-cpp","vllm","cherry-studio","localai","exo","sglang","mlx"],"stacks":[],"scenarios":["switch-models","run-models-locally","use-ai-compliantly-in-china","learn-ai"],"problems":[],"reasonZh":null,"sources":["editorial","github-pool"],"record":"https://aimap.page/api/v1/software/deepinfer.json","pages":{"en":"https://aimap.page/en/software/deepinfer/","zh-hans":"https://aimap.page/zh-hans/software/deepinfer/"}}