DeepSeek V4.1 Flash hands-on tests and comparisons

A user combined WorkBuddy, REA and DeepSeek-V4.1-Flash-1017 to download WeChat Channels videos, noting free overseas access until Oct 31, roughly 100M tokens per day, and a high risk of account bans.

Two community quants of DeepSeek V4.1 Flash on TP=2 DGX Sparks scored 316 vs 320 on 356 questions (a tie), decoded at ~86 tok/s, but differed by 3.2 GB per Spark, giving 170K vs 2.1M token KV cache.

A user reports V4.1 Flash rose to second on a health Q&A leaderboard, beating GPT models and Opus 5.5 and trailing Sonnet 5.5 by just over a point, at $0.005 per question—about 1/16 of Sonnet's cost.

A user reports that DeepSeek V4.1 Flash 1017 is now free to use in OpenCode, without further details.

Fingerprint matched official 1017, but on OpenDesign about 98% of streamed thinking was empty, with 6 dropped streams, 10 error events, 11 HTTP 5xx/524s, and 140–150s median per task versus 30–47s official.

OpenRouter's Oct 5–11 data shows Chinese models at ~48% of requests with DeepSeek first at 24.5%, citing V4.1 Flash's DeepSWE 77.0 and $0.15 per-million input pricing.

OrcaRouter's own infrastructure data shows 1017 cuts time-to-first-token by about 30% versus 0910, with shorter reasoning traces, lower cost per task, and Terminal-Bench rising from 31.2 to 42.3.

The author tested DeepSeek Harness 0.2.1-rc.2 with DeepSeek V4.1 Flash via API, MiniMax V3.1 Flash and GLM-5.3 Flash, sharing screenshots of the run.

A short reaction noting the recent DeepSeek V4.1 Flash update scores 45 on the Artificial Analysis Intelligence Index, tied for second among open models, with no further testing.

The UD-IQ1_S quant targets 128GB devices like DGX Spark, matching the original model 82.8–85.1% of the time, with a 110.6GB Engram table on SSD and a custom llama.cpp PR build required.

The poster says the 1017 build is close to Astra on multimodal generation, beats it in some video cases, and costs a tenth or less, while dismissing fingerprint-based distillation claims.

The user notes that free tiers usually train on inputs, but DeepSeek V4.1 Flash in OpenCode has zero data retention and isn't used for training, which they find notable.