DeepSeek-V3

Headlines · 3

1. 3d
   Huawei, HKUST BalanceRoute boosts decode throughput ~15% in LLM serving tests

   Machine Heart (Jiqizhixin) reports that a team from HKUST led by Zhou Zijie and Huawei 2012 Labs' theory department proposed BalanceRoute for data-parallel load imbalance in LLM serving, where synchronized decoding leaves lighter groups waiting for the busiest. The method models this waiting as an imbalance objective and turns short-term load prediction into online routing (BR-H), with requests pinned to a worker after assignment. In the paper's 96-NPU Ascend 910C experiment running DeepSeek-V3 671B, the authors report about 15% higher decode throughput than the strongest baseline and a 1.20 ms median scheduling time; they say the work was accepted at NeurIPS 2026 and deployed in MathBot-Pie. The figures are author-reported and have not been independently reproduced.

   机器之心 · 🔥 0
2. 4d
   vLLM adds day-0 support for Reflection's 501B Beam model with Blackwell optimizations

   vLLM published day-0 serving support for Reflection's Beam, a 501B-parameter mixture-of-experts model with 23B active parameters per token, optimized for NVIDIA Blackwell. The engine handles reasoning, tool calling and structured outputs out of the box and can run Beam with tensor, data or expert parallelism, with a Docker image and launch command provided. The post says coding agents autonomously tuned the implementation so batch-1 decode is 2.71× faster and prefill 1.87× faster than the first version, and that Beam's DFlash2 drafter reaches 708 tokens/s for a single request on TP8 with an acceptance length of 3.7. It also describes day-0 reinforcement-learning support built with the SkyRL team, reporting AIME25 average accuracy rising from about 61% to about 79% after 50 training steps.

   vLLM Blog · 🔥 0
3. Oct 12
   AirLLM claims 2.8T-parameter Kimi K3 runs on a single 4GB GPU, adds low-VRAM training

   AirLLM is an open-source inference and training library that streams weights one layer at a time, keeping only a single layer resident on the GPU to cut VRAM use sharply. According to its repository, it runs 70B models on a single 4GB card without quantization, distillation or pruning, and reports running Kimi K3 (2.8T) in 3.72GB, Qwen3.8-Flash-Next (125B) in about 5.95GB and DeepSeek-V3 (671B) in roughly 12GB. Recent updates also add small-VRAM training, fine-tuning Qwen3.8-Flash-Next (125B) under 6GB and Qwen3.8-27B at about 2GB. All VRAM and training figures are the project's own reported measurements and have not been independently verified.

   GitHub Trending（每日） · 🔥 0

Experience and discussion from the community

Share my DeepSeek-V3 experienceAsk about DeepSeek-V3