DeepSeek V4.1 Flash hands-on tests and comparisons

The author estimates ~760GB to load the 8-bit 763B model, more than two 128GB DGX Sparks, with 4-bit still ~380GB, and lists four benchmark wins over Opus 5 against seven losses.

Given a free-form prompt to build an impressive game, the model produced a complete street-racing demo with procedural roads, drift-charge mechanics, five weather stages and WebAudio synth music in 65 minutes for about ¥2.5, far outperforming the author's earlier Gemini test.

Five hands-on complaints: hallucinations still need a second model to check; fewer tokens doesn't mean faster; quality depends heavily on the harness; multi-file refactors need a stronger planner; Search+Fetch starts charging per call Nov 1.

Testing V4.1-Flash-1017 on LisanBench, the poster found a chain of 64 two-letter words and concluded DeepSeek almost certainly RL-trained on that benchmark.

A short video post arguing that DeepSeek's new V4.1 Flash does not outperform Astra 6.

A reply urges buying a zcode subscription and using the V4.1 Flash API, calling ChatGPT and Claude a scam on value without giving numbers.

The open release scores 77.0 on DeepSWE, above Astra's 74.1 and Opus's 74.2 and close to Argon's 77.9, with per-task cost equal to or below GPT-6 Luna and strong 3D modeling.

On the same voxel prompt Opus 5 produced the better result, while DeepSeek V4.1 finished faster at roughly 4% of the cost, prompting a 24x price question.

A short reaction post noting DeepSeek V4.1 Flash 1017 weights are freely downloadable and cheap to use, questioning the prospects of closed models without offering any tests or data.

A user had V4.1 Flash independently handle editing, music, motion and pacing for an MG animation in DeepSeek Harness, costing ¥1.86 paid from free credits.

Lucebox released ROCMFP GGUF quantizations on day zero, runnable on 128GB Strix/Gorgon Halo machines with only 6GB offloading, and Lucebox users get optimized quants automatically.

Split-screen floating-island scenes show GPT with denser detail but DeepSeek structurally complete, with the author citing off-peak pricing, free search APIs until Oct 31, and doubting the $0.0044 figure.