DeepSeek V4.1 Flash hands-on tests and comparisons

Roundup of this month's API price cuts, including DeepSeek V4.1 Flash going from 13 to 29 OpenRouter providers, plus the point that cache reads are 96–99% of coding-agent token spend.

A user says DeepSeek V4.1 Flash 1017 is currently their main model on two DGX Sparks and asks what models the other person plans to run.

The author deployed DeepSeek V4.1 Flash 1017 on four DGX Sparks with two self-made recipes, SGLang and TensorFold, and the attached image shows the comparison results.

Per-million output token prices: DeepSeek V4.1 Flash at $0.60–$1.20 versus GPT-5.4 Mini at $4.50 and GPT-5.4 at $15, making Flash up to 25x cheaper.

The author ran Cohere's prefill method 100 times on DeepSeek-V4.1 (1017) with zero supplied URLs; the model's self-generated hosts were mostly legitimate threat-intel sources, and 169.254.169.254 only appeared when explicitly asked to read cloud metadata.

Single-needle retrieval hit 16 of 16, but at 125K with thinking off DeepSeek scored 0 of 12 on two-hop questions, recovering to 12 of 12 when split or with thinking on; GLM took 48.8s to first token at 1M.

GPT-6.1 Terra scores 48 on the Artificial Analysis intelligence index, only 3 points above DeepSeek V4.1 Flash at 45, while Flash's peak output price of $1.2 is about a quarter of Terra's; Opus 5.5 leads at 58.

Two TensorFold music machines built by DeepSeek-V4.1-Flash-1017 on four DGX Sparks: Mia's run took 93.9 min and 272k tokens for a 9/10, bertholomus' 91.9 min and 327k tokens for 8/10.

In OpenCode, DeepSeek V4.1 Flash 1017 read a real open-source Rich issue, reproduced the bug, fixed it and wrote a test, going green in about 30 seconds.

A hands-on test ran DeepSeek-V4.1-Flash-1017 via DeepSeek Harness 0.2.1-rc.2 with 1M-token context and the TensorFold recipe across four DGX Sparks, producing a working browser music machine in one try in 93.9 minutes and 272k tokens, scoring 9.0/10 on a browser check at about 70 tok/s decode.

In a timed exploit challenge only frontier models finished: Claude Mythos 5.1 retrieved the flag in 3h28m for about $100 and GPT-5.6 Sol took 6h10m, while DeepSeek V4.1 Flash and Kimi K3.1 Pro both stalled at identifying the memory leak.

DeepSeek's API now redirects V4 Flash Vision Exp calls to V4.1 Flash 1017; a llama.cpp fix made with the model on a 192GB M2 Ultra was merged, logging 28.1 minutes at 215 tok/s prompt and 30.8 tok/s output, with the smallest GGUF at 98.6GB.