DeepSeek-V4.1

Headlines · 3

1. 7h
   SGLang v0.5.22 ships new model support and up to 2.3x throughput gains

   SGLang has released v0.5.22, with its release notes stating the version covers 950 pull requests from 245 contributors. The release adds support for models including Clef / Clef Flash, PPLX-Decider v1 / v1.1, EmbeddingGemma 2, LLaDA2.2, HunyuanImage-3.0-Instruct and Kandinsky 6. The notes say DeepSeek-V4.1 gains up to 2.3x more throughput per GPU on H200 and now runs on AMD MI350X / MI355X with up to 1.9x throughput from DSpark, while GLM-5.3 can combine decode context parallelism with MTP for up to 2.1x faster decoding. The release also requires Python 3.11 or newer and deprecates options such as --enable-dp-attention, making it a breaking-change upgrade.

   SGLang Releases · 🔥 8
2. 5d
   vLLM: bitwise-consistent train-inference RL for faster training

   A vLLM blog post describes making reinforcement-learning training and inference numerically bitwise-consistent: the training forward pass calls vLLM's inference kernels directly, and those high-performance kernels are modified so rollout and training produce an identical logprob for every sampled token (diff = 0). The authors report that on DeepSeek-V4-Flash the high-performance baseline's train-inference KL rose from 2e-3 to 29.9 while reward fell back, whereas their consistent run held diff at 0 across 100 training steps at 91.3% of the baseline's rollout throughput; carried over to Nemotron-3.5 (a Mamba-2 hybrid with NVFP4), the consistent run kept its SWE-bench Pro score around 0.6 while the baseline dropped from 0.56 to about 0.22. The process is packaged as an agent skill and released with recipes and code, and DeepSeek-V4.1 has already reached 0 diff while performance work is being finalized.

   vLLM Blog · 🔥 0
3. Oct 10
   ByteDance Seed finds periodic blind spots in DeepSeek-V4 long-context retrieval

   A ByteDance Seed team with researchers from Princeton, Stanford and UC Berkeley systematically tested DeepSeek-V4 and V4.1 and describes a "phase sensitivity" effect in the paper "Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression": because CSA compresses the KV cache with overlapping windows of length 8 and stride 4, shifting the same information by a few tokens can sharply change whether it is retrieved. In a 128K needle-in-a-haystack test, the base V4-Flash model scored 69.92% in the best position group versus 29.69% in the worst, a 40.23-point gap; the post-trained V4-Flash-0731 narrowed this to 19.14 points, while V4-Pro showed 34.77 and 14.84 points for base and post-trained versions. The article says DeepSeek had already cut the compression stride from 4 to 2 in V4.1-Flash and added an Engram module, shrinking the gap to 6.1 points without removing the periodicity.

   36氪 人工智能 · 🔥 0

Experience and discussion from the community

Share my DeepSeek-V4.1 experienceAsk about DeepSeek-V4.1