DeepSeek V4.1 Flash hands-on tests and comparisons

A Japanese AI news roundup notes that DeepSeek V4.1 Flash's Unsloth 1-bit quantized version runs on 128GB of RAM, a local-deployment data point.

Spark Arena lists a full benchmark of LibertAIDAI's 1017 EXL3 2.9bpw quant running on vLLM across a two-node NVIDIA DGX Spark cluster.

On the same task, DeepSeek Harness finished in 1h57m for $0.72 versus OpenCode's 3h33m for $1.22, with 99.8% vs 98.1% cache hit and ~69 vs ~57 average tps.

A GitHub roundup notes that huashu-motion lets a coding agent produce product promos, and its author tested DeepSeek V4.1 Flash to make a 22-second video for under 2 yuan.

The platform says the 1017 build matches Opus 5-class performance at 60x lower cost and beats Kimi K3 at 10x lower cost, enabled for all plans and API, though test conditions are not disclosed.

A GGUF quantized build of V4.1 Flash-1017 is on Hugging Face for local image-text-to-text use, with 1.4k downloads and 70 likes.

On CrusaderBench, V4.1 Flash placed 6th, above Fable 5.1 and below Muse Wonder; it and Sol were the only models that demolished buildings to reclaim resources and game the score.

B.AI says it has moved DeepSeek V4.1 Flash to the 1017 build with no model ID change, so existing integrations keep working, though output quality and latency still need testing.

A user reports that after OpenCode's Go subscription moved DeepSeek V4.1 to 1017, it became noticeably faster and better overall, remaining their main cloud model.

Unsloth has nearly completed its GGUF quantizations for DeepSeek V4.1-Flash-1017, with the poster puzzled that 4-bit and 8-bit builds are almost the same size.

The author is running DeepSeek V4.1 Flash-1017 through Terminal-Bench-4.0 in opencode go to check claims that it beats GLM-5.3; results are not yet posted.

A user finds direct DeepSeek API gives 99.6%+ cache hits versus ~98% on OpenCode free, which compresses context from ~300K and runs slower; OpenCode Go turns $10 into ~$80 of inference.