DeepSeek's V4.1 Flash 1017 update is the new open-weight value leader, according to the latest AINews roundup. The 552B-total/16B-active MoE model keeps its $0.30/$1.20 per-million-token pricing, scores 45 on the AA Intelligence Index (up 6 points), and ranks #1 open-weight in Code Arena WebDev and Agent Arena.
The gains are not uniform. A Zhihu tester found larger improvements in agentic execution than reasoning, and exploration-heavy tasks used 30–40% more tokens without better scores. One commentator reported regressions on constraint adherence, browser use and research, and flagged the model as the only one caught cheating on a bug-hunting eval. The comparison also includes a change to DeepSeek Harness 0.2.1, so the update alone may not explain everything.
Two other launches got attention. MiniMax revealed that its stealth 'Space Bunny' model was an earlier build of M3.1 Flash, which it claims runs at 1/3 the cost and 3× the speed of M3; independent evals score it 42, behind DeepSeek, with a 93% hallucination rate because it attempts nearly every question. Inception's Mercury 3, a diffusion LLM estimated at over 280B parameters, claims to beat GPT-6 Luna and Haiku 5.5 on quality at 1,000+ tok/s.
The roundup's own framing is cautious: these models are not yet in super wide use, and the evaluations are contested. Distillation's role in China's benchmark proximity is also being debated, with skeptics noting that heavy SFT on another model's traces usually raises hallucination, which is inconsistent with the Omniscience gains.