DeepSeek's Flash-1017 Gains 11.1 Points on Terminal-Bench 4.0 in Its Own Table, and the Card Gives No Release Date
A.I. / news
DeepSeek's Flash-1017 Gains 11.1 Points on Terminal-Bench 4.0 in Its Own Table, and the Card Gives No Release Date
The MIT-licensed checkpoint is now what the deepseek-flash API name serves, but active parameters, training data and hardware needs are all absent from the card.
DeepSeek's Hugging Face card for DeepSeek-V4.1-Flash-1017 lists gains over the earlier V4.1-Flash on all 11 benchmarks it reports, the largest being 11.1 points on Terminal-Bench 4.0. Every figure is supplied by DeepSeek. The card, which calls the release a "minor update", gives no release date, no training-data statement and no active-parameter count.
DeepSeek's API documentation says the deepseek-flash model name now points to DeepSeek-V4.1-Flash-1017. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but are retired, and their requests are served by the new checkpoint at the Flash price, the documentation said.
What the 1017 card claims against V4.1-Flash
The model page puts the new checkpoint at 42.3 on Terminal-Bench 4.0, against 31.2 for V4.1-Flash, using the Claude Code harness. On DeepSWE v1.1, run with the Mini-SWE-V2 harness, it reports 77.0 against 74.2. Humanity's Last Exam without tools moves from 36.8 to 42.6, or from 39.1 to 44.0 on the text-only subset.
- Terminal-Bench 4.011.1 points
- ProgramBench8.8 points
- Apex Bench v1.17.5 points
- HLE without tools5.8 points
- Agents' Last Exam3.6 points
- DeepSWE v1.12.8 points
- ZeroBench-main2 points
Source: DeepSeek-V4.1-Flash-1017 model card on Hugging Face, accessed 2026-10-19
The smallest gains are on the visual tests, such as ZeroBench-main at 51.0 against 49.0. The card describes the update as improving reasoning, coding, agentic tasks and vision, and as shortening task execution time and output length. It publishes no figure for either of the last two.
| Benchmark | Flash-1017 | V4.1-Flash |
|---|---|---|
| Terminal-Bench 4.0 | 42.3 | 31.2 |
| ProgramBench | 29.1 | 20.3 |
| DeepSWE v1.1 | 77.0 | 74.2 |
| NL2Repo-Bench | 68.2 | 65.4 |
| BabyVision (with tools) | 92.5 | 89.6 |
The V4.1-Flash column matches the figures The Next Web reported from DeepSeek's September launch material: 74.2 on DeepSWE v1.1, 36.8 on Humanity's Last Exam and 20.3 on ProgramBench. The comparison is therefore DeepSeek against its own earlier numbers, not a third party's rerun. The Next Web said it had not verified them.
Size, licence and hardware
The Hugging Face listing shows 763B parameters in Safetensors metadata. The V4.1-Flash card says the language backbone is 552B and that a 196B conditional-memory component called Engram accounts for much of the difference. It gives 8B active parameters per token when reading a prompt and 16B when generating. The 1017 card does not restate either figure.
The licence is MIT for both the code and the weights. That puts it in a different position from Meta's Muse Spark open weights, which carry a $20 million resale threshold. Context is 1M tokens, and the API documentation allows up to 384K output tokens.
The card lists no GPU requirement. Its SGLang Docker example sets --shm-size 32g, which is a shared-memory flag and not a memory minimum. Tensors are published in BF16, F32, F8_E4M3 and I8. For a smaller open model from another lab, see MiniCPM-V 4.7.
What the API costs
DeepSeek's pricing page lists deepseek-flash at $0.15 per million input tokens on a cache miss, $0.003 on a cache hit and $0.60 per million output tokens off-peak. Peak rates are double. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays.
The same page still lists a deepseek-v4-pro column, version DeepSeek-V4-Pro-0813, at $0.66 input and $1.98 output off-peak. The Next Web reported on September 10 that all V4-Pro requests would be routed to V4.1-Flash and billed at Flash rates from 04:00 UTC on September 14. The two accounts conflict, and the pricing page does not say which applies.
The version suffixes 1017 and 0813 are not defined as dates on the pricing page. The ZeroBench, Chartography and BabyVision scores use tool access, and the card does not say which tools.
Per-request billing for DeepSeek's Search and Fetch tools, priced at $0.75 per 1,000 successful requests, begins at 00:00 Beijing time on November 1, 2026, the pricing page said. They are free until then when used through DeepSeek Harness, Claude Code or Codex.
Sources
More in A.I.
- 01Qwen3-4B-Deslop Cuts Its Slop Score From 121 to 12.5 on a MacBook Air, and Two of 40 Answers Came Out DamagedThe Show HN fine-tune by Erik Markus Kannike uses a rank-16 LoRA and a 175-byte neuron edit; its blog gives no licence, its Hugging Face card says Apache-2.0.
- 02HybridAgent Paper Claims 10.4% Gain and 2.7x to 3.7x Lower Cost, but Its Own Text Gives 10.4% Two Different MeaningsThe Stanford, UIUC and Together AI preprint ships MIT-licensed 8B and 14B agents, and no independent run of its numbers has been reported.
- 03Gemini 4 Argon Reaches the API at $2 and $10 Per Million Tokens, With a No. 1 Ranking and No ScoreGoogle's October 13 post says Argon tops the Vals Index but publishes no figure for it, and the introductory price doubles on November 30.
- 04DeepSeek Swapped V4.1-Flash-1017 Behind Its API Name, and the New Card Lists NL2Repo at 65.4 Where the Old One Said 64.0The MIT-licensed update calls itself minor and is now what deepseek-flash serves, with no active-parameter count and no release date on the card.