DeepSeek-V4.1-Flash-1017 Lifts Terminal-Bench 4.0 From 31.2 to 42.3 Under the Same MIT Licence
A.I. / news
DeepSeek-V4.1-Flash-1017 Lifts Terminal-Bench 4.0 From 31.2 to 42.3 Under the Same MIT Licence
The 'minor update' on Hugging Face beats Claude Opus 5 on two of five vendor-run rows and its card lists two different parameter counts.
DeepSeek has posted DeepSeek-V4.1-Flash-1017 to Hugging Face under the MIT licence, and its model card reports Terminal-Bench 4.0 at 42.3, up from 31.2 for the V4.1-Flash it replaces. The card gives no release date, only the 1017 suffix.
The card calls the checkpoint "a minor update to DeepSeek-V4.1-Flash" with improvements in reasoning, coding, agentic tasks and visual benchmarks. It also says the model "reduces task execution time and output length". The context window is 1 million tokens and the page shows no access gate.
Vendor-run scores against the previous Flash
All figures below are DeepSeek's own. The card says the Claude Code harness was used for Terminal-Bench 4.0 and the Mini-SWE-V2 harness for DeepSWE v1.1.
| Benchmark | 1017 | V4.1-Flash | Claude Opus 5 |
|---|---|---|---|
| Terminal-Bench 4.0 | 42.3 | 31.2 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 74.0 |
| ProgramBench | 29.1 | 20.3 | 37.0 |
| HLE, no tools | 42.6 | 36.8 | 56.3 |
| Apex Bench v1.1 | 72.1 | 64.6 | 65.8 |
The card also lists GPT-5.6 Sol, GLM-5.3 and a model it labels K3. On DeepSWE v1.1 the 1017 build scores 77.0 against 73.0 for Sol and 74.0 for Opus 5, and on Apex Bench v1.1 it scores 72.1 against 54.3 for Sol and 65.8 for Opus 5. Those are the two rows where it leads both US models.
- Claude Opus 551.8 points
- DeepSeek-V4.1-Flash-101742.3 points
- GPT-5.6 Sol39.9 points
- GLM-5.337.9 points
- DeepSeek-V4.1-Flash31.2 points
- K312.6 points
Source: DeepSeek-V4.1-Flash-1017 model card, accessed 2026-10-17
Opus 5 still leads on the other three rows above. The gap on Humanity's Last Exam without tools is 13.7 points.
Two parameter counts on one card
The earlier V4.1-Flash card describes "552B backbone parameters" and 8 billion active per token during prefill and 16 billion during decode. Its Safetensors metadata lists 763 billion parameters, and the page does not explain the difference. The 1017 card lists the same 763 billion in its metadata and gives no active-parameter count.
The Next Web reported the original V4.1-Flash launch on Sept. 10 at 552 billion total parameters. Whether the 1017 build changes the architecture is not given, and DeepSeek calls it a minor update.
Availability and what the page leaves out
The card says no Inference Provider deploys the model. Serving examples use vLLM and SGLang, and the card recommends a temperature of 1.0, top_p of 1.0 and max_tokens of at least 256,000.
DeepSeek has not published API pricing for the 1017 build on the pages reviewed. For the September model, The Next Web, by reporter Ana Maria Constantin, put off-peak output at $0.60 per million tokens and said V4-Pro requests route to V4.1-Flash from 04:00 UTC on Sept. 14 until a V4.1-Pro arrives. DeepSeek has given no date for that model.
For context on how vendor tables compare with outside tests, see Mercor's map of work hours no benchmark covers, and for another open-weights release, Reflection's 501B Apache-licensed model.
DeepSeek has not said whether a technical report will accompany the checkpoint.
Sources
More in A.I.
- 01Vals Test: Three of Six AI Agents Threatened Court Over a £350 Train RefundTold to try everything to recover a refund the rules no longer allowed, Grok 4.7 sent 187 messages in 15 simulated days while Opus 5.5 sent eight in twelve weeks.
- 02MiniCPM-V 4.7 Is Apache 2.0 With 3B Active Parameters and No Benchmark TablesOpenBMB's first mixture-of-experts vision model claims a win over Qwen3.6 35B-A3B in prose and charts, and 13 tokens a second on an iPhone, but the model card prints no scores.
- 03Gemini 4 Argon Opens to the API at $2 per Million Tokens, and Loses 5 of 19 Rows in Google's Own TableGoogle's October 13 launch post leads with a Vals Index ranking; the comparison table behind it shows Argon behind GPT-6 Astra or Opus 5.5 on both Terminal-Bench tests and on OSWorld 2.0.
- 04StarSkirmish Has Nine Models Write StarCraft Bots in One Hour, and the Best Wins 7% Against Top Human BotsGemini 4 Argon lands in a three-way tie at 37 to 39, but it ran at a lower reasoning setting than most rivals and its runs vary the most.