DeepSeek V4.1 Flash hands-on tests and comparisons

Third-party platform NanoGPT now serves V4.1 Flash 1017 with image input, tool calling and 1M-token context, offering standard and Thinking modes at the standard 1x subscription rate.

The official release says the new version handles sparse-prompt "wish-style" coding requests, generating runnable projects from minimal instructions and references.

A Persian-language post says DeepSeek V4.1 Flash 1017 costs roughly 60x less than Opus 5 while scoring about 3 points higher on the DeepSWE v1.1 benchmark.

A Persian post saying V4.1 Flash 1017 beats GPT 5.6 Sol on Terminal-Bench while costing roughly 17 times less, without specific scores.

A user generated a Grand Budapest Hotel-style video of four seasons passing outside a train window with V4.1 Flash 1017 and called the result very good.

The author frames the update as a win for vibe coders and indie hackers: notable jumps in terminal-task benchmarks, faster inference and unchanged dirt-cheap pricing, with a breakdown thread.

Hands-on complaint: even at 150 tok/s, the model spends too long thinking before doing anything, which hurts a Flash-tier model.

Within seven hours of release, an NVFP4 quant appeared on Hugging Face, shrinking the model from 475 GiB to 400 GiB with bit-exact routed experts.

WorkBuddy Intl now supports the 1017 model free until October 31, citing faster runs and shorter outputs but no benchmarks or post-free pricing.

After the DeepSeek V4.1 Flash 1017 release, a user tags @ollama hoping local-run support arrives soon.

Open weights under MIT; scores 77.0 on DeepSWE vs Opus 5's 74.0 at about $0.20 per task (~50x cheaper), but still trails on Terminal-Bench and ProgramBench.

A comparison claim that the low-cost DeepSeek V4.1 Flash outperforms Claude Opus 5 on a coding test, with no test details given.