DeepSeek V4.1-Flash (763B, 8B prefill/16B decode) introduces a causal encoder-decoder split that cuts KV cache 8x and costs $0.27/task vs V4 Pro's $0.67.
September 12, 2026
brightray analysis
Summary
DeepSeek V4.1-Flash uses a novel causal encoder-decoder architecture with separate active parameter budgets for prefill (8B) and decode (16B) stages, achieving ~1% sparsity and up to 8x KV cache reduction versus V4 Flash. Independent benchmarks place it at AA Index 40 and AutomationBench-AA 69% — above V4 Pro despite being far cheaper — with MIT licensing and 1M-token context. The release signals a strategic pivot toward inference-optimized, agent-friendly architecture and reinforces DeepSeek's view that data quality now dominates post-training algorithm research.
Why it matters
- The causal encoder-decoder split (8B prefill / 16B decode) reduces KV cache footprint up to 8x versus V4 Flash, making long-context agentic workloads dramatically cheaper to serve.
- Cost-adjusted benchmarks show $0.27/Intelligence Index task — roughly 7x cheaper than GLM-5.3 and Kimi K3 and 2.5x cheaper than V4 Pro — while matching or beating them on AutomationBench-AA (69%) and AA-LCR (84%).
- Local inference on consumer hardware is now viable: 300+ TPS on 4-GPU setups with <32GB RAM via SSD offloading, and confirmed operation on a 128GB M5 Max, lowering the barrier for self-hosted frontier models.
- DeepSeek's post-training notes say data quality ROI now exceeds novel algorithm ROI, aligning with Prof. Jie Tang and signaling a broader industry shift in where the gains come from.
- Critics note high verbosity (89k tokens/task, 25% above GLM-5.3) and observed brittleness in multi-agent team mode, where solo V4.1 outperformed coordinated sub-agent setups in some workflows.
Community notes—
No notes yet — be the first.
See every signal in the Feed