DeepSeek V4 lands, Hygon adapts Day-0, Huawei pairing confirmed
Busy day. V4 drops with coordinated silicon support, the Anthropic-shaped pivot becomes undeniable, and a pure-inference GPU play crosses unicorn status.
DeepSeek V4 released, open weights, Huawei chip partnership confirmed
V4 ships as open weights with an explicit Huawei Ascend pairing called out at launch. The framing: breaking the closed-frontier monopoly, with domestic silicon as the deployment substrate rather than an afterthought. Hygon's DCU completed Day-0 adaptation in parallel (see below), so two of the main domestic accelerator lines are usable from hour zero.
Open weights plus coordinated Chinese hardware readiness is the exact axis Kir-News tracks.
Read at QbitAI → https://www.qbitai.com/2026/04/406359.html
Xiwang becomes first Chinese inference-only GPU unicorn, S3 chip targets 90% token-cost cut
Xiwang has raised ~4B RMB across seven rounds at a >10B RMB valuation, all on the bet that inference, not training, is where the volume goes. The S3 claims ~99% GEMM and ~98% Flash Attention utilization, native FP4, up to ~600GB LPDDR6 (largest among domestic GPUs), and PCIe Gen6. Target price point: ¥0.01 per million tokens. CUDA compatibility is maintained at 99%+ via custom low-level shims. CEO Wang Zhan argues 2026 inference demand will run 4–5x training demand.
Inference economics and domestic silicon with real CUDA compat, directly relevant to self-hosting math.
Read at QbitAI → https://www.qbitai.com/2026/04/406036.html
Chinese labs' new benchmark is Anthropic, not OpenAI
Zhipu, MiniMax and Moonshot are now explicitly racing Claude Opus 4.6 on agentic coding. Zhipu's API platform hit 1.7B RMB ARR in 2025, up 60x year-over-year, and pushed prices 83% higher in Q1 2026 while call volume kept growing, which is about as clean a pricing-power signal as you get. MiniMax M2.7 ran 100+ rounds of scaffold self-optimization for ~30% eval gains. Moonshot booked more revenue in the 20 days after K2.5 than in all of 2025.
The shift from chatbot benchmarks to agentic API revenue is the actual industry turn, and the numbers now back it.
Read at Recode China AI → https://recodechinaai.substack.com/p/forget-openai-chinas-ai-labs-are
JiuwenClaw Team Skills: a reusable multi-agent SOP package format
Huawei-backed openJiuwen published a spec that packages an entire agent team (roles, task split, conflict resolution) into a single folder, the Team Skill. A generator tool produces them, a hub hosts them, and the demo validated zero-adaptation execution on both Claude Code and Cursor. The medical triage demo dynamically assembled 23 specialist agents per case.
Concrete agent-composition primitive that's framework-portable, worth watching as a pattern.
Read at QbitAI → https://www.qbitai.com/2026/04/406393.html
Hygon DCU completes Day-0 adaptation for DeepSeek V4
Hygon's DCU accelerator line was ready for V4 on release day, mirroring the Huawei pairing in the launch announcement. No benchmarks or architectural notes in the flash, but the coordination itself is the story: release and domestic hardware enablement are now synchronized events.
Signals that domestic-chip deployment lag for frontier Chinese models is approaching zero.
Read at 36Kr AI → https://36kr.com/newsflashes/3780507701072904?f=rss
Skim pile
- Qianli Tech's ASD 4.0 in 460k vehicles, in-car Super Eva agent on STEP 3.5 Flash · QbitAI · auto-domain embodied-agent stack claiming Huawei ADS parity · 55
- GPT-5.5 launches, 82.7% on Terminal-Bench 2.0, Codex writes its own load balancer · QbitAI · US context but useful reference point on agent coding frontier · 52
- Feishu Project goes "AI Friendly" with MCP server, CLI, and AAMP async agent protocol · QbitAI · enterprise agent plumbing, API calls 5M to 23M daily in a year · 44