DeepSeek V4.1-Flash-1017 Posts Gains on All 11 Listed Benchmarks, Every One Of Them Its Own
A.I. / news
DeepSeek V4.1-Flash-1017 Posts Gains on All 11 Listed Benchmarks, Every One Of Them Its Own
The October 17 update ships 763B-parameter weights under MIT with a 1M-token window, but the model card gives no active-parameter count, hardware guidance or training data.
DeepSeek released V4.1-Flash-1017 on October 17, an update to its V4.1-Flash model, and published the weights on Hugging Face under the MIT licence. The company's model card calls it "a minor update" and lists higher scores on all 11 benchmarks in its table.
Every figure in that table is DeepSeek's own. No independent party is named as having reproduced any of them as of Sunday afternoon.
What the model card says about V4.1-Flash-1017
The model card reports gains in reasoning, coding, agentic tasks and visual benchmarks, and says the update reduces task execution time and output length. TechNode reported the same release date and the same MIT licence.
The Hugging Face metadata lists 763B parameters in BF16, F32, F8_E4M3 and I8 tensors. The context window is 1M tokens. The card recommends a max_tokens setting of at least 256K, temperature 1.0 and top_p 1.0.
The weights are public. The page shows a file listing and download counts with no access-request form.
Where the largest gains are
The biggest jump is on Terminal-Bench 4.0, from 31.2 to 42.3. The card says that test ran on the Claude Code harness, not DeepSeek's own. DeepSWE v1.1 ran on the Mini-SWE-V2 harness and moved from 74.2 to 77.0.
- Terminal-Bench 4.011.1 points
- ProgramBench8.8 points
- Apex Bench v1.17.5 points
- HLE without tools5.8 points
- Automation-Bench v1.0.65.7 points
- Agents' Last Exam3.6 points
- DeepSWE v1.12.8 points
- NL2Repo-Bench2.8 points
Source: DeepSeek-V4.1-Flash-1017 model card on Hugging Face, accessed 2026-10-25; differences computed by The Terminal
The card also compares the model with K3, GLM-5.3, GPT-5.6 Sol and Opus 5. The Terminal covered Kimi K3 in a Cerebras hardware walkthrough.
| Benchmark | V4.1-Flash | V4.1-Flash-1017 |
|---|---|---|
| Terminal-Bench 4.0 | 31.2 | 42.3 |
| DeepSWE v1.1 | 74.2 | 77.0 |
| ProgramBench | 20.3 | 29.1 |
| HLE without tools | 36.8 | 42.6 |
| ZeroBench-main (with tools) | 49.0 | 51.0 |
What the card does not say
The card does not give an active-parameter count, so it is not possible to tell from the page how much compute each token uses. It names no hardware requirement for running 763B parameters. It carries no statement about training data.
The release date is also missing from the card, which gives only the year 2026. The October 17 date comes from TechNode.
Google's own Gemini 4 Argon post reports 77.9% on DeepSWE v1.1. DeepSeek's figure on the same test is 77.0. Each number comes from its maker, and the Google post as read does not state which harness produced its score, so the two should not be set side by side as a ranking.
API changes alongside the weights
DeepSeek's API documentation says the hosted service is now served by DeepSeek-V4.1-Flash-1017 and that "the calling method remains unchanged." The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but the documentation says the corresponding models "have been retired."
The same page announces a Search / Fetch API and a developer preview of DeepSeek Harness. The documentation page read here does not state prices for either.
For comparison, Block's Dessa speech model launched with no benchmarks at all. DeepSeek at least publishes a table, and the next test is whether anyone outside the company reproduces it.
Sources
More in A.I.
- 01Eka's 1,500 Picks an Hour Is a Lab Number, and Its Own Investors Say SoThe Cambridge, Mass. startup raised $150 million for a vision-force-action robot model. Its headline throughput comes from induction and bin-picking tests, and customer pilots are unnamed.
- 02Cerebras Says Kimi K3 Fits on 59 Wafers, and Gives No Price for Any of ThemA vendor walkthrough puts Moonshot's 2.8-trillion-parameter model at 35 plus 24 wafer-scale chips and claims 1,000 to 2,000 tokens a second. Every input is an assumption.
- 03Apollo's Torsten Slok Says AI Has Paid Off Only for the Magnificent 7. His Note Prints No FiguresThe Oct. 24 note says the S&P 493 shows flat margins and no productivity gain. Ramp's September index has 55.9 percent of US businesses paying for AI.
- 04ChatGPT Ads Reach 89 Countries as OpenAI Says More Than 100,000 Advertisers Have Run CampaignsOpenAI added 26 countries on October 21. Ads stay limited to the Free and Go plans, and the company gave no per-country prices or user counts.