DeepSeek ships V4.1 Flash: 552B MoE with CED asymmetric architecture, beats V4 Pro
APPSO · wechat · 2026-09-10
DeepSeek has officially released V4.1 Flash across web, app, and API — a 552B-parameter MoE model with native multimodal vision that optimizes for capability, speed, and throughput rather than raw scale.
The tech report highlights a new CausalEncoderDecoder (CED) asymmetric architecture: 8B parameters activate at input, 16B at output, cutting inference cost while retaining capacity. Combined with CompressedSparseAttention2 (cross-layer KVCache reuse + Top-K indexing in Full/Reindex/Reuse modes), FP4 KV Cache, and DSpark speculative generation, HBM needs drop to 1/4 and SSD to 1/8 versus the previous generation — KVCache is 437x smaller than the first-gen model.
With RL post-training on top, V4.1 Flash beats flagships including V4 Pro on agent benchmarks. DeepSeek is sunsetting V4 Pro (requests auto-route to Flash pricing) and cut API prices, keeping off-peak half-price billing. Harness v0.1.5 ships alongside: update system prompts without losing KVCache, parent-child agent two-way communication, and experimental AgentTeams multi-agent collaboration. Tencent WorkBuddy, CodeBuddy, and OpenCode already integrate the model.
More from coding & agent
- New article: 'Long live AGENTS.md' on repo-level instructions for coding agents — _jaydeepkarale · 2026-09-19
- Dev's take: Codex worth the $200 tier, Claude fine at $100, top-tier Cursor underused — lxfater · 2026-09-19
- AgentSky launches as an 'OpenRouter for agents': run 40+ coding agents in-browser and compare costs side by side — Aiden_Tech_Ai · 2026-09-19
- qwen-code SDK v0.1.13 fixes microcompaction to preserve prompt-cache reuse in long agent sessions — github-actions[bot] · 2026-09-19
- Agent engineering splits into three tiers: scripts, System One decision models, and reasoning models — arpit_bhayani · 2026-09-19
- Lovable migrates its internal sandbox from Vite to OJ, a Rust-based solution — w_hgm · 2026-09-19