DeepSeek V4.1 Flash hands-on tests and comparisons

On 105 planted bugs across two real repos, DeepSeek V4.1 Flash 1017 scored 42.7, ahead of Muse Wonder 42, Opus 5.5 41.7 and Kimi K3.1 Pro 36.

After repeated video-generation attempts the author finds results far worse than Opus 5.5, which raises their opinion of deepseek-v4.1-flash-1017; more tests planned.

A creator uses DeepSeek-V4.1-Flash-1017 to produce a PS1-style Tears of the Kingdom demake, saying the whole thing cost about $2.5.

Vals AI finds DeepSeek V4.1 Flash on Fireworks costs a third less than DeepSeek's API yet generates at 103 tok/s versus 286, so price does not predict speed.

A 20-task, two-round Terminal-Bench 2.0 test of DeepSeek-V4.1-Flash-1017 found Claude Code and Hermes at 75% and DeepSeek Harness at 70%, while Claude Code used ~40% fewer input tokens and 18 tool turns vs 29.

The author teases a new benchmark comparing Kimi K3.1 Pro, Minimax M3.1 Flash, DeepSeek V4.1 Flash-1017 and Muse Spark Open, but shares no results or numbers yet.

A developer used the official API in the DeepSeek Harness macOS app to one-shot a voxel pagoda garden with day/night cycles and falling petals, burning 29M tokens at 230 tok/s.

A reply advises that with enough memory any quant runs, but smaller is faster when nothing fits fully in VRAM, recommending Maya-XS for DeepSeek V4.1 Flash-1017.

Per-million-token pricing for Kimi K3.1 Pro versus DeepSeek V4.1 Flash 1017, where off-peak cache hits at $0.003 are about 67x cheaper than $0.2.

An EXL3 2.5bpw quant (~185GB) with TensorFold TP2 runs 1M context on two DGX Sparks at ~107 tok/s on code and ~2.4K tok/s prefill, but tool calls degrade most at 2.5bpw, so 4 Sparks at 4.0bpw is still recommended for agents.

Of 129,280 vocab rows in the open weights, 1,249 are reserved, including control tags like <political> that alone trigger a refusal; 799 blank rows remain to be swept.

A security researcher ran DeepSeek V4.1 Flash 1017 fully locally in VulnHunter, locating all 24 planted vulnerabilities from source and producing working proofs for 22 of them.