DeepSeek V4.1 Flash: 552B params at 420-507 tok/s, matches Claude Opus-5 on agent benchmarks
APPSO · wechat · 2026-09-12
DeepSeek released V4.1 Flash, unifying its chat, vision and expert models into one entry point. Despite doubling backbone params to 552B, it runs at 420-507 tok/s thanks to a redesigned architecture: a Causal Encoder-Decoder (CED) split that nearly halves long-input prefill compute, CSA2 cross-layer KV sharing with FP4 quantization, SWA Bounded Replay cutting persistent KV cache to 1/8 of the prior gen, and a 196B lookup-only Engram memory module. Per-token global KV cache is just 890 bytes — 1/437 of DeepSeek-V1. On agent benchmarks it hits 74.2% on DeepSWE v1.1 and beats Claude Opus-5 and GPT-5.6 Sol on Terminal-Bench 2.1 and AutomationBench, though gaps remain on the hardest tasks (Terminal-Bench 4.0: 31.2 vs Opus-5's 51.8). Post-training focused on task data engineering rather than new RL algorithms, and the report shows agent framework differences swing the same model's scores from 74.2 to 65.6.
Related event: DeepSeek Launches V4.1-Flash with Native Vision and 1M Context(5 posts)→
More from coding & agent
- Training a messaging agent for format adherence with dual-ring protocol before RL — cephaloform · 2026-09-12
- A 70-char character-level tokenizer for brainfuck coding agents lands on Hugging Face — cephaloform · 2026-09-12
- Dev Ditches Claude Code for Codex: 'No More Found the Smoking Gun' — TheMoonMidas · 2026-09-12
- The benchmarking problem: we're evaluating adapter-harness-model triplets, not models — teortaxesTex · 2026-09-12
- Git AI Duo Joins OpenAI as Company Pledges to Keep Agent-Tracking Tool Open Source — Dimillian · 2026-09-12
- Tasking an AI agent to buy 4 Tokyo concert tickets: the 'Math Olympiad' of assistants — annbordetsky · 2026-09-12