DeepSeek V4.1 Flash hands-on tests and comparisons

An analysis says V4.1 Flash 1017 cuts KV cache to 890 bytes per token via CSA2 and FP4, scoring 77.0% on DeepSWE v1.1 versus 74.0% for Opus 5 and 42.3% on Terminal-Bench 4.0 in Claude Code.

A user runs V4.1 Flash 1017 fully in NVFP4 on one RTX Pro 6000 MaxQ and two ASUS GX10 DGX Sparks, reaching 54 tok/s with more tuning to come.

A user says their AI bug bounty automation keeps producing good findings with DeepSeek V4.1 Flash 1017 and plans to open-source the whole setup soon.

The post reports V4.1 Flash 1017 scoring 77.0% on DeepSWE v1.1 vs Opus 5's 74.0%, Terminal-Bench 4.0 rising 31.2% to 42.3% in Claude Code, and open kernels for SM100 and Ascend 950.

SGLang v0.5.22 ships 950 PRs, including up to 2.3x output token throughput for DeepSeek-V4.1-Flash on Hopper and up to 10x faster weight loading via InstantTensor.

A user screenshots DeepSeek V4.1 Flash 1017 running at 244 tokens per second inside DeepSeek Harness, with no task details.

Using official figures of 155T tokens per day and $38 per GPU-hour, the author estimates ~2,370 GPUs, $2.16M daily revenue and 75–84% GPU-only gross margin, noting the assumptions.

Comparison of three 45+ index models: V4.1 Flash 1017 at 206 tok/s, index 45, $0.26; Muse Wonder 145 tok/s, index 52, $4.52; Sonnet 5.5 136 tok/s, index 56, $5.46.

After using DeepSeek V4.1 Flash as a daily model for 19.5 hours and switching back to GLM 5.3 Flash, the same 4 boxes and 8 chats ran at 177.2 tok/s for GLM versus 140.9 tok/s for DeepSeek, though DeepSeek's builds looked nicer.

A side-by-side video comparing outputs from DeepSeek V4.1 Flash and Claude Opus 5.5, both generated through the same YouArt MCP setup.

A user ranks DeepSeek V4.1 Flash 1017 first among local AI models, ahead of GLM 5.3 Flash and Qwen 3.8 Next Flash, without citing test data.

Timing on the same task: DeepSeek V4.1 Flash 1017 took about 1h30m, Claude Opus 5.5 1h19m33s, and GPT-6.1 Sol just 26m57s.