DeepSeek Swapped V4.1-Flash-1017 Behind Its API Name, and the New Card Lists NL2Repo at 65.4 Where the Old One Said 64.0
A.I. / news
DeepSeek Swapped V4.1-Flash-1017 Behind Its API Name, and the New Card Lists NL2Repo at 65.4 Where the Old One Said 64.0
The 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.
DeepSeek has replaced the model behind its deepseek-flash API name with DeepSeek-V4.1-Flash-1017, and its own docs say the calling method is unchanged. The Hugging Face card calls the checkpoint "a minor update to DeepSeek-V4.1-Flash" and lists gains on all 11 benchmarks it compares.
The API documentation says requests to the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but those models are retired and the requests are now served by the 1017 checkpoint. The page gives no date for the switch and no new prices.
What the 1017 card says
The model card states an MIT licence and a recommended context window of 1M tokens, with max_tokens of 256K or more. The Hugging Face safetensors listing shows 763B parameters.
The card does not state active parameters, hardware requirements, training data or a release date. The earlier V4.1-Flash card describes a multimodal mixture-of-experts model with 552B backbone parameters, 8B active per token during prefill and 16B during decode,. That card also shows 763B in the safetensors listing, so the 552B and 763B figures coexist for the same family and the 1017 card does not reconcile them.
The weights are not gated on the pages as fetched.
Old checkpoint against new, vendor-supplied
Every figure below is from DeepSeek's own card, not an independent run.
| Benchmark | V4.1-Flash | 1017 |
|---|---|---|
| HLE, no tools | 36.8 | 42.6 |
| Terminal-Bench 4.0 | 31.2 | 42.3 |
| DeepSWE v1.1 | 74.2 | 77.0 |
| ProgramBench | 20.3 | 29.1 |
| NL2Repo-Bench | 65.4 | 68.2 |
| AutomationBench v1.0.6 | 54.8 | 60.5 |
| Agent's Last Exam | 31.8 | 35.4 |
- V4.1-Flash31.2 points
- V4.1-Flash-101742.3 points
Source: DeepSeek Hugging Face model cards, accessed 2026-10-19
The largest jump is Terminal-Bench 4.0, up 11.1 points. The smallest of the rows above is DeepSWE v1.1, up 2.8.
One row does not reconcile. The 1017 card puts the old model at 65.4 on NL2Repo-Bench, while the old model's own card lists 64.0. The 1017 card also shows the old model at 64.6 on "Apex Bench v1.1", where the old card reports 65.6 on MathArena Apex, so those two cannot be matched. DeepSeek has not said whether the benchmarks were re-run with newer harness versions.
The card says the update also reduces task execution time and output length, which it says lowers cost. It gives no token counts.
Why the name matters
DeepSeek formally released V4.1 Flash on September 10, according to TechNode, citing a notice to API users. From September 14, V4 Pro requests were routed to V4.1 Flash and billed at Flash rates. DeepSeek said Flash had surpassed V4 Pro in performance, cost, speed and total completion time, and the TechNode article carries no test data for that.
A developer who never changed a model string had Pro requests rerouted on September 14 and has since had the Flash model replaced. Anyone who pinned a benchmark baseline to September behaviour is comparing against a checkpoint that no longer answers.
TechNode reported off-peak prices of RMB 1 per million input tokens on a cache miss and RMB 4 per million output tokens, with peak prices twice as high. The 1017 docs page lists no prices, so those are the September figures.
The card's comparison table also sets the checkpoint against Opus 5, GPT-5.6 Sol, K3 and GLM-5.3. Opus 5 scores higher on most of those rows. DeepSeek leads several, including DeepSWE v1.1 and AutomationBench.
On licences, Meta's Muse Spark Open sits behind an access form, and Liquid's LFM2.5-24B-A2B carries a revenue cap. DeepSeek's card carries neither.
DeepSeek has not published a changelog for 1017 or said when it will update the Models and Pricing page.
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.
- 04Liquid's LFM2.5-24B-A2B Licence Ends at $10 Million in Revenue, and Its Table Scores Qwen Below Qwen's Own CardThe 24B mixture-of-experts model's weights are downloadable, but the commercial terms sit on a separate licence page that the model card does not summarise.