2026-10-18
DeepSeek-V4.1-Flash-1017 raises the scores on the same machines
On 17 October 2026 DeepSeek published a weights-only update. Same architecture and memory as the 9 September launch, with higher coding, agent and vision scores and shorter answers.
On 17 October 2026 DeepSeek published DeepSeek-V4.1-Flash-1017. It is the weight refresh of DeepSeek-V4.1-Flash, released on 9 September. The SophyAI catalog now serves this checkpoint.
What stays the same
Architecture, configuration, tokenizer and prompt format are those of the launch. The release still has 48 shards, about 510 GB, mixed MXFP4/MXFP8 precision and the official minimum of 614 GB of HBM. Native context stays at 1 million tokens. The license stays MIT.
The catalog machines therefore stay the same: vLLM on 8× H200, on 8× B200/B300 and, as an advanced profile, on 4× B200. The image vllm/vllm-openai:deepseekv41-flash-0909 loads the new weights, because the configuration is identical. The Mac path with ds4 Metal stays on the Q2 quant of the September launch.
Where the scores go up
The numbers in the table are the ones DeepSeek published on the 1017 model card. They are not SophyAI measurements.
| Benchmark | 9 September | 1017 |
|---|---|---|
| HLE without 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 |
| Automation-Bench | 54.8 | 60.5 |
| Agents' Last Exam | 31.8 | 35.4 |
| Apex Bench | 64.6 | 72.1 |
| Chartography, with tools | 78.9 | 83.0 |
| BabyVision, with tools | 89.6 | 92.5 |
| ZeroBench-main, with tools | 49.0 | 51.0 |
The widest jumps are Terminal-Bench 4.0, from 31.2 to 42.3, and ProgramBench, from 20.3 to 29.1. HLE without tools moves from 36.8 to 42.6. On agent tasks, Automation-Bench rises from 54.8 to 60.5 and Agents' Last Exam from 31.8 to 35.4. Vision with tools also gains on Chartography, BabyVision and ZeroBench.
DeepSeek says that, along with the scores, the model writes shorter outputs and finishes tasks sooner. An independent measurement by sference, run on 17 October on the same problems and the same stack, finds an AIME median of 1,124 output tokens against 1,681 for the 9 September weights, about 21% less, with 100 correct answers out of 120 against 101. At reasoning_effort 1 the median falls to 790 tokens. These numbers are not SophyAI measurements either.
In the catalog
The served checkpoint is deepseek-ai/DeepSeek-V4.1-Flash-1017. The product profile stays vLLM on 8× H200. Hourly prices in the catalog are still those from the Vast search of 12 September 2026: we have not remeasured them on the new weights, and we do not publish our own tok/s.