DeepSeek V4.1 Flash-1017 Lands, and the Local Crowd Races to Run It

DeepSeek released V4.1 Flash-1017 on October 17. @MiaAI_lab announced it was live that morning, saying it “looks like it improved on everything.” @kimmonismus called it “roughly on par with Opus 5” and “a significant jump compared to DeepSeek 4.1 Flash regular,” adding that it outperforms GLM 5.3 across the board and even GPT-5.6 Sol.

On the benchmarks DeepSeek reported, @kimmonismus says the new Flash beats GPT-5.6 Sol and Opus 5 on selected coding and agent tests. On DeepSWE it scores 77.0, up from 74.2 for the previous Flash, against 73.0 for GPT-5.6 Sol and 74.0 for Opus 5. On Automation-Bench it scores 60.5, against 54.8 for the previous Flash, 45.8 for GPT-5.6 Sol and 50.3 for Opus 5. Terminal-Bench 4.0 is the exception: Flash-1017 scores 42.3, ahead of the previous Flash (31.2) and GPT-5.6 Sol (39.9) but behind Opus 5 at 51.8.

Price is a big part of the story. Per million uncached tokens, @kimmonismus lists DeepSeek Flash-1017 at $0.15 input and $0.60 output off-peak ($0.30 / $1.20 at peak), compared with $4 / $20 for GPT-5.6 Sol and $5 / $25 for Claude Opus 5. He notes that GPT-6 Luna is cheaper per uncached token at $0.10 / $0.50, but that Flash is substantially cheaper than Sol and Opus while beating them on some of DeepSeek’s reported tests. DeepSeek also reports faster task completion and shorter outputs. The weights are open under the MIT license, with a 1M-token context window.

The local-hardware community moved quickly. @MiaAI_lab ran V4.1 Flash on three and four DGX Sparks using the original quant DeepSeek posted, with no further quantization. On 4x Sparks she measured 87 tokens per second on prose in a single stream and 163 tokens per second across four concurrent streams. She also noted that the earlier vLLM recipe for V4.1 Flash “should” work with the new release, and that she is working on a TensorFold recipe.

Her early verdict is measured. In a creativity test, she asked the model to produce 100 HTML files that would “look stunning,” with zero repetitive designs. It “produced good results, but it’s not the BEST.” She called GLM 5.3 Flash still “one of the best models you can run comfortably on 2x DGX Sparks,” and said she plans a direct comparison but is not yet confident the new DeepSeek is the right choice for a two-Spark setup.