DeepSeek V4

Headlines · 2

1. 23h
   iFlytek to unveil Ascend 950-trained flagship LLM at 1024 event

   iFlytek plans to announce a new flagship general-purpose large model trained on Huawei's Ascend 950 at its October 24 1024 Developer Festival, a bid to escape the domestic-compute bottleneck. The article says Spark has been the only mainstream LLM trained entirely on Chinese chips for three years, but Ascend 910B's memory and bandwidth left its long-context and complex-code training about 5x less efficient than Nvidia's H200; Ascend 950's larger memory and bandwidth raise theoretical efficiency against H200 to 65%–85%, giving Spark a base for top-tier models. Commercially, the piece notes H1 2026 revenue of RMB 11.623 billion, up 6.52% year on year, a RMB 204 million net loss attributable to shareholders, a 2.65% decline in government-business revenue, and chairman Liu Qingfeng's August admission that iFlytek was visibly absent from the code and agent boom. The event is also expected to cover funding progress for the spun-out Ciyuan Xinghuo unit and chip and supernode plans.

   36氪 人工智能 · 🔥 6
2. Oct 10
   ByteDance paper finds periodic blind spots in DeepSeek V4 chunked KV-cache compression

   A ByteDance team paper, “Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression,” says the chunked KV-cache compression used by the DeepSeek V4 series creates “phase sensitivity”: adding only decorative equals signs ahead of a code snippet flipped DeepSeek-V4-Flash-Base's completion between 8 and 32, with an error rate of about 71.3% when the count of equals signs modulo 4 was 0 or 1 and about 91.5% accuracy when it was 2 or 3. In a 128K-token needle-in-a-haystack test with roughly 16,000 key-value pairs, moving a single pair changed accuracy by up to 40.2% for V4-Flash-Base and 34.8% for V4-Pro-Base, with gaps of 19.1% and 14.8% remaining after post-training; halving the compression stride in DeepSeek-V4.1-Flash narrowed the gap to 6.1% but did not remove it. Control models trained from scratch on the Qwen3-0.6B architecture showed the same periodic swings whenever fixed-length chunked compression was used, while a full-attention baseline did not, which the authors present as a structural flaw rather than a tuning issue.

   雷峰网 AI · 🔥 0

Experience and discussion from the community

Share my DeepSeek V4 experienceAsk about DeepSeek V4