DeepSeek V4.1 Flash hands-on tests and comparisons

The author is running the Tool eval bench hard mode on DeepSeek V4.1-Flash-1017 in a sparkline dashboard and will share scores after multiple runs.

A user reports DeepSeek V4.1 (off-peak) outperforms Terra high at similar cost with faster API and better price-performance, while Terra max barely matches Sol medium at nearly 2x cost per task and runs at 25 tps in Codex.

The author says GPT-6.1 Terra is fast, but DeepSeek V4.1 Flash is faster, cheaper and genuinely open source, which OpenAI should have done.

A user reports DeepSeek V4.1 Flash running at 50 tok/s on an M3 Mac Studio using oMLX.

The author argues US chip export bans pushed Chinese labs toward efficiency, noting DeepSeek V4.1 Flash 1017 ships the same weights it serves via API and runs on four home DGX Spark units.

inference.net releases deepseek-v4.1-flash-1017-cyber, the first of several cyber models, served at cost in US data centers with zero data retention by default and KYC required.

The open-source huashu-motion project lets Claude Code build launch films with auto-scored music; the author says a 22-second film on DeepSeek V4.1 Flash cost just $0.27.

A user says DeepSeek V4.1 Flash performed okay but rambled instead of getting to the point, noting Opus has spoiled them.

A post notes that Cursor paid-plan users can now select DeepSeek V4.1 Flash (1017), with no further details.

On a multi-turn adversarial debate benchmark, V4.1 Flash 1017 (high) rises from 1577 to 1593 (+16) while using 29% fewer reasoning tokens on average; Kimi K3.1 Pro drops from 1691 to 1603.

Strata 0.2.6 preloading CPU-side experts into RAM lifts single R9700 code decode from 24.5 to 27.6 tok/s and 4x R9700 from 77 to 84.7 tok/s, at the cost of 55s instead of 34s load time.

A two-week, 348-commit agent pipeline has Claude Code orchestrate, DeepSeek v4.1 Flash implement and GLM v5.3 verify, all running on a $300 used mini PC.