Search “Reasoning”

DeepSeek open-sources the inference engine DeepInfer on GitHub

On October 29, 2026, DeepSeek open-sourced the inference engine DeepInfer on GitHub; this engine is the inference engine behind DeepSeek-V4.1-Flash. Leiphone reported the same day that DeepInfer includes technologies such as MoE expert parallelism, asynchronous batch scheduling, custom kernels, and speculative decoding, and disclosed that a single cluster needs to process approximately 155 trillion tokens per day. Its Zero-Overhead Batch Scheduler lets the CPU schedule computation one round in advance, overlapping with GPU forward passes, and handles cross-round state misalignment through total_inflight_tokens accounting, Remap mapping, and DeferredRelease deferred reclamation. Currently, only offline inference capabilities are open; components such as online KV Cache storage and cross-replica load balancing have not yet been fully disclosed, and the official team has not separately published end-to-end speedup data for this scheduling mechanism relative to synchronous solutions.

vLLM Blog Publishes Technical Retrospective on Kimi K3 Inference Optimization

On October 16, 2026, the vLLM blog published a technical retrospective titled "Hyper Optimizing Kimi K3 in vLLM: Moving Fast and Responsibly." According to a TMTPost·AGI report, NVIDIA CentML engineer Benjamin Chislett used a Megakernel written in 12,000 lines of CuTe DSL to optimize inference for Moonshot AI's Kimi K3; the vLLM blog reports over 840 tokens/s at batch 1 and a 1.5x to 2x improvement in decoding throughput across multiple configurations, while SemiAnalysis estimates 423 tok/s/user on GB300 NVL72, a 1.68x improvement. The vLLM blog states that agent-optimized inference code is now outpacing human review and maintenance capacity, and the team therefore proposed a fast-track model optimization RFC so that reviewers judge designs from high-level abstractions rather than reviewing code line by line. The report also mentions that NVIDIA is developing optimizations for other models in vLLM, such as DeepSeek V4.1.

vLLM Blog Publishes Technical Retrospective on Kimi K3 Inference Optimization

On October 16, 2026, the vLLM blog published a technical retrospective titled "Hyper Optimizing Kimi K3 in vLLM: Moving Fast and Responsibly." According to a TMTPost·AGI report, NVIDIA CentML engineer Benjamin Chislett used a Megakernel written in 12,000 lines of CuTe DSL to optimize inference for Moonshot AI's Kimi K3; the vLLM blog reports over 840 tokens/s at batch 1 and a 1.5x to 2x improvement in decoding throughput across multiple configurations, while SemiAnalysis estimates 423 tok/s/user on GB300 NVL72, a 1.68x improvement. The vLLM blog states that agent-optimized inference code is now outpacing human review and maintenance capacity, and the team therefore proposed a fast-track model optimization RFC so that reviewers judge designs from high-level abstractions rather than reviewing code line by line. The report also mentions that NVIDIA is developing optimizations for other models in vLLM, such as DeepSeek V4.1.

View story:vLLM Blog Publishes Technical Retrospective on Kimi K3 Inference Optimization

DeepSeek open-sources inference framework DeepInfer

On October 20, 2026, DeepSeek open-sourced its internal inference framework DeepInfer. According to Yuchen Jin, citing data disclosed by DeepSeek, a single cluster processes 155 trillion tokens per day and generates $38 in revenue per GPU-hour (off-peak pricing); DeepInfer supports both NVIDIA Blackwell and Huawei Ascend. A subsequent 36Kr report stated that its accompanying open-source deepseek-recipe supports NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. On October 21, swyx, citing Latent.Space data, said a single cluster processes 155T tokens per day with a 96% input cache hit rate, and argued that the lifetime value of each sold GPU has risen by about 30%. On October 24, Nathan Lambert said DeepSeek's single cluster processes 150T tokens per day, exceeding the combined open-source model inference volume of OpenRouter, Fireworks, Together, and BaseTen, and noted that the actual number of clusters is unknown; on the same day, LMSYS published a same-hardware benchmark comparison, stating that SGLang meets or exceeds DeepInfer on DeepInfer's own benchmark, with prefill throughput up to 1.04x and decode step length shortened by 1.1%–7.4% on 64x GB300.

DeepSeek open-sources inference framework DeepInfer

On October 20, 2026, DeepSeek announced the open-sourcing of its internal production-grade inference framework, DeepInfer. According to Latent Space, the framework powers DeepSeek-V4.1-Flash, is customized based on vLLM, supports distributed inference, speculative decoding, SWA Bounded Replay, and structured output, and targets the NVIDIA SM100 and Huawei Ascend 950 platforms; the published single-cluster production figures are 155 trillion tokens processed per day, a 96% input cache hit rate, 400 ms first-token latency, and a generation speed of 230 tokens/s. The report also noted that components such as KV cache management and load balancing have not yet been fully open-sourced. The next day, DeepSeek published the DeepInfer technical report on Zhihu, disclosing internal details including the hardware selection for Prefill and Decode and the KVCache targets, and providing data showing a single cluster serving 155 trillion Tokens/day at $38 per card per hour at off-peak prices. The framework is currently open-sourced, but the KV cache system, Prefill/Decode disaggregation, and load balancing required for a complete service are not fully included in this open-source project.