DeepSeek Open-Sources V4.1 Flash: 552B Asymmetric MoE Replaces V4 Pro
机器之心 · wechat · 2026-09-10
DeepSeek V4.1 Flash is now live on web, app, and API, and open-sourced. It's a 552B-parameter MoE built on a new Causal-Encoder-Decoder architecture with asymmetric activation—only 8B active on input, 16B on output—plus new pretraining methods and larger-scale RL post-training. Officially, it surpasses V4 Pro across core metrics.
Benchmarks: Top scores in DeepSWE v1.1 (74.2, slightly ahead of Opus 5), CyberGym, and Automation-Bench, but Terminal-Bench 3.0 (30.0) trails Opus 5 (43.3) and GPT-5.6 Sol (34.4); strong on coding/agent tasks but not uniformly ahead.
Cost: The asymmetric design targets agent workloads with tens of thousands of input tokens but few output tokens. KV Cache shrinks dramatically—HBM needs down to 1/4, SSD to 1/8—so weight reads become the bottleneck, explaining the measured 400 token/s with no context-length slowdown.
Pricing/migration: V4 Pro API retires Sept 14, requests auto-routed to V4.1 Flash at off-peak prices of ¥0.02 (cache hit input), ¥1 (miss), ¥4 (output) per million tokens; peak is 2x. Consumer web/app modes are merged onto V4.1 Flash with deep thinking + search. DeepSeek Harness v0.1.5 ships alongside with file uploads and plugin UI extensions; Tencent WorkBuddy, CodeBuddy, and OpenCode have integrated it.
More from Infra
- MiniMax H3 runs fully on a Galaxy S26: 5-second video with audio in ~7 minutes — BidFunny4862 · 2026-10-27
- Unconventional AI acquires inference chip firm NeuReality in 1,000x efficiency bid — NaveenGRao · 2026-10-27
- Musk: robot RAM demand will dwarf everything else, and 100% of superintelligence compute goes to deep space — elonmusk · 2026-10-27
- RTX 5090 hits $7K as PC gaming community goes into 'denial' over AI boom — PAstynome · 2026-10-27
- Wan2GP lands on Pinokio: 1-click AI video on 6GB GPUs — cocktailpeanut · 2026-10-27
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27