DeepSeek V4.1 Flash hands-on tests and comparisons

The author breaks down Artificial Analysis Coding Agent Index 2.0: Codex + DeepSeek V4.1 Flash scores 47.3 for $0.39, versus GPT-6.1 Sol at 54.1/$1.50, GPT-6 Astra at 54.7/$7.04 and Opus 5.5 at 61.7/$11.42.

The experimental build keeps the sparse-attention indexer, shared query projection and compressor in BF16 for only +0.15GB, matching the original model's token selection 100% past 16K, though full-model accuracy stays ~73% and the 210GB 2.9bpw repo leaves little KV cache room across two DGX Sparks.

Kimi K3.1 Pro generated a fully procedural Three.js lighthouse with PBR materials and an interactive camera, but the author finds its 3D ability close to yet below DeepSeek V4.1 Flash 1017.

A poster makes a bold pre-test claim that DeepSeek V4.1 Flash 1017 is stronger than Kimi K3.1 Pro, saying they hope to be proven wrong once they run their own tests.

A user spots a community GSQ-RCO quantized build of DeepSeek V4.1 Flash 1017, noting Strata has no official support yet and PR3589 is unmerged.

Kimi K3.1 Pro jumps from 54.0 to 81.6 on BuildingBench, essentially tying DeepSeek V4.1 Flash 1017 at 81.9, at an estimated $12.58 per task versus DeepSeek's $3.01.

Runs the ~340GB Maya-S quant on a Windows box with one RTX 5090 and 256GB RAM at about 31 tok/s (up to 35); disabling --mmap-experts lifted it from 7.7 to 25 tok/s, while AVX512 gave no speedup.

The author compares Kimi K3.1 Pro and DeepSeek V4.1 Flash pricing: K3.1 Pro is 32–35% cheaper than K3 but still 12–16x the price of V4.1 Flash, with cache reads 61x apart.

A developer wrote one new test for the DeepSeek-V4.1 engine's Engram reader and it surfaced two bugs before ever passing.

In a one-shot Poland summer tour build, DeepSeek V4.1 Flash 1017 cost $0.79 in 5h30m, versus Opus 5.5 at $94.06 in 7h10m, Gemini 4 Argon at ~$5.20 and MiniMax M3.1 Flash at ~$2.58.

A user comparing Kimi-K3.1-pro with DeepSeek V4.1 Flash 1017 says the gains are not profound, DeepSeek Flash remains stronger at 3D, and Kimi does not fix the areas where DeepSeek fails.

In Menlo's Asimov humanoid robot simulation, DeepSeek V4.1 Flash (in DSH) was faster while Kimi K3.1 was slower but more precise.