DeepSeek-V4.1-Flash: 8B Activated Model Beats 1.6T V4-Pro, Local Deployment Tested

DeepSeek-V4.1-Flash is a newly open-sourced model that the author argues deserves a larger version number due to its architectural advances and benchmark results.

Architecture and Scale

DeepSeek-V4.1-Flash uses a new CED architecture with native vision support. Its total parameter count is 552B, but the key innovation is only 8B/16B activated parameters, enabling faster inference and higher throughput. The model weights are stored in FP8 precision, totaling approximately 510GB. Some MoE routing experts use FP4, leaving little room for further quantization.

Benchmark Performance

Two benchmark charts released by DeepSeek show V4.1-Flash surpassing K3, GLM-5.3, Opus5, and GPT5.6-Sol across coding benchmarks. More strikingly, it comprehensively beats the previous flagship V4-Pro (1.6T parameters). DeepSeek's API documentation confirms V4.1-Flash has fully surpassed V4-Pro, so all deepseek-v4-pro requests are now routed to V4.1-Flash at the lower Flash price.

Compared to GPT-6 Astra, a gap remains on most benchmarks. The author suspects Astra's AutomationBench score uses a different private test set. However, on DeepSWE v1.1 — which tests an AI's ability to understand an existing codebase, make changes, and pass functional acceptance — V4.1-Flash nearly matches Astra.

Local Deployment Requirements

Due to the 510GB FP8 weight size, full-context local deployment is estimated to require an 8×H200 or 16×PPU server (budget ~300k RMB). Even with compression, 8×96GB RTX PRO 6000 GPUs are needed. Quantization headroom is minimal because weights are already FP8/FP4 and KV cache is already 1/4 of previous generation.

The author notes that for most practical coding tasks, GLM-5.3-Flash is a more cost-effective local choice. MiaAI reportedly ran GLM-5.3-Flash on 2×DGX Spark (~70k RMB) with moderate speed/quality, and later ran V4.1-Flash on 3×DGX Spark. Q2 GGUF quantizations have also appeared. The author personally still uses Qwen3-32B and Qwen3.8-27B on 2×4090 for many workloads.

Best Practices and Harness Testing

DeepSeek's technical report includes a controlled-variable experiment across six harness families (8 configurations): Claude Code, Codex, OpenCode, Pi, mini-SWE, and DeepSeek Harness (Minimal/Standard/PTC modes). All ran at Max reasoning effort (=100) on DeepSWE v1.1 and Terminal-Bench v2.1.

Key findings:

DeepSeek Harness Minimal mode scored highest on both benchmarks, suggesting more complex frameworks don't automatically yield higher success rates.

Claude Code or Codex with V4.1-Flash achieve near-peak performance out of the box.

The highest-scoring combination isn't necessarily best for daily use; the author dislikes DSH and recommends personal testing.

The author criticizes uncontrolled "pelican riding a bicycle" tests circulating online and demonstrates a simple knowledge question ("What model is 桂纶镁 -5.3-Flash?") where V4.1-Flash and Kimi K3 fail without web search, while GPT-5.6 Sol succeeds.

Real-World Usage

The author values V4.1-Flash's speed and low API cost for long, token-heavy tasks like video downloading skills and video-to-HTML conversion, reporting high satisfaction.

UI Update

DeepSeek's website now shows a simplified UI with only "deepseek-flash" model selection, aligning with the API routing change.