DeepSeek V4.1 Flash hands-on tests and comparisons

The author releases an IQ2_XXS GGUF quant and recipe, reporting ~21 tok/s on fresh prompts and 25 on repeats on a single RTX 5090, versus ~7 with MXFP4 experts.

The poster asks for firsthand experience using DeepSeek V4.1 Flash 1017 with Claude Code; the post itself contains no results.

Argues that DeepSeek-V4.1-Flash scoring above larger models shows scaling is not the only path, with data quality and architecture mattering too.

A user gave the same prompt to DeepSeek V4.1 Flash and Gemini 4 Argon to generate a Tesla car flyover scene, joking that game prices should come down.

The lead item summarizes the official update: web/app/API sync, higher benchmarks, shorter runtimes and outputs, plus DeepSeek Harness and CodeBuddy support.

The post cites MIT-licensed weights, $1.20 per million output tokens at peak, and an official claim of beating Opus 5 on DeepSWE while questioning self-reported scores.

A custom EXL3 quant of V4.1 Flash 1017 (TP=2, 2.80BPW, 203GB, 2.4M cache) scores 95.17% code, 94.93% science and 87.64% tool call vs the full model, 5% faster on short replies and 5% slower on long ones.

Roundup of the official changelog and HF weights, scores like HLE 42.6, coding leaderboard results, pelican SVG style shifts and ~300 tok/s observations.

Given the same prompt for a racing game with minimap, timer and AI rivals, DeepSeek ran on the first try but looked plain, while Gemini 4 delivered more polish.

A joking post notes that DeepSeek V4.1 Flash 1017 is being benchmarked against Opus 5.5 rather than Claude Haiku, calling the tiering confusing.

Across 125K–1M needle-in-a-haystack prompts on a single GB300, DeepSeek scored 4/6, 4/6, 5/6, 4/6 versus GLM's 6/6, 6/6, 6/6, 4/6; all misses were two-hop chains and 1M TTFT was ~49–51 s.

A blogger ran the same 3D floating-island web scene prompt on V4.1-Flash-1017 and GPT-6.1 Sol; GPT was more detailed, DeepSeek still delivered a usable scene at $0.0044 per API call.