5 articles
Index of editorial tags and filtered coverage.
On 30 September 2026 DeepSeek announced on WeChat, per Reuters and 1 October coverage, a partnership with Huawei to develop and open-source programming infrastructure for Ascend chips, including the TileLang language and compute/communication libraries. The stated goal: an independent GPU software ecosystem that reduces reliance on Nvidia.
Announced ~10 September 2026, DeepSeek-V4.1-Flash is a native multimodal (vision+text) 552B MoE with an asymmetric causal encoder–decoder (8B active on input, 16B on output), sharply smaller KV cache, peak/off-peak pricing, and retirement of V4-Flash—direct pressure on closed labs’ cost per token.
Per Bloomberg, TrendForce, and The Decoder reports (Sep 2026), Liang Wenfeng is driving a cluster of at least 160,000 Ascend 950DT chips at an Inner Mongolia datacenter, with deliveries spanning Q4 2026–Q1 2027, to train toward 2-trillion then 8-trillion-parameter models—a gradual migration under export controls, not a clean break from Nvidia.
The DeepSeek Elastic Compute (DSec) technical report, arXiv/HF 2609.22978 (~18 Sep, late-Sep coverage), describes a production platform with FnCall, container, microVM, and full-VM backends: ~160 nodes per unit, ~3 million sandboxes/day, 380,000+ concurrent, and 5,000+ creations/sec—co-designed with RL to mitigate reward hacking.
DeepSeek released competitive reasoning models with reportedly low training costs, altering market expectations.