DeepSeek-V4.1-Flash-1017
DeepSeek·
LLMsopen-weightcloud + localFrontier
Overview
October checkpoint of DeepSeek-V4.1-Flash, served under the same deepseek-flash API name. Same Causal Encoder-Decoder architecture and size as the September release (552B backbone, 8B active during prefill and 16B during decode); the Hugging Face card calls it a minor update, and coverage reports the config unchanged with new weight shards. 1M-token context, image and text input, MIT license, open weights. Launched alongside DeepSeek's first public Search and Fetch APIs.
Capabilities and innovations
552B backbone MoE, 8B active (prefill) / 16B active (decode)Native multimodal image and text input1M token context windowMIT license, open weightsRe-post-trained checkpoint on the V4.1-Flash architectureShips with DeepSeek's first public Search and Fetch APIs
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| Humanity’s Last Exam (no tools) | 42.6% | vendor HLE 42.6 (no tools), up from 36.8 on the September build. DeepSeek's own launch figure, read from press coverage of the launch table (MarkTechPost, The Decoder): the Hugging Face model card and DeepSeek's site were not reachable from the environment this entry was written in, so the figure should be re-checked against the card. One article claims the model leads HLE; its own table puts Claude Opus 5 at 56.3, so that claim is not carried. |
| Terminal-Bench 4.0 | 42.3% | vendor Terminal-Bench 4.0 42.3, up from 31.2. DeepSeek's own launch figure, read from press coverage of the launch table (MarkTechPost, The Decoder): the Hugging Face model card and DeepSeek's site were not reachable from the environment this entry was written in, so the figure should be re-checked against the card. Measured in DeepSeek's own harness, the same caveat the September entry records; its own generation on its own scale. |
| DeepSWE v1.1 | 77% | vendor DeepSWE v1.1 77.0, up from 74.2. DeepSeek's own launch figure, read from press coverage of the launch table (MarkTechPost, The Decoder): the Hugging Face model card and DeepSeek's site were not reachable from the environment this entry was written in, so the figure should be re-checked against the card. DeepSeek harness rather than DeepSWE's fixed mini-swe-agent setup. Vendor-baseline check: the V4.1-Flash column in the 1017 table (74.2 DeepSWE, 31.2 Terminal-Bench 4.0, 36.8 HLE) matches what DeepSeek published for that build in September, so the deltas are not a restated baseline. |
- GPQA Diamond: not reported by the vendor — No GPQA figure in the 1017 launch coverage. The September build's 90.9 is not carried over.
- SWE-bench Verified: not reported by the vendor
- SWE-bench Pro: not reported by the vendor — DeepSeek reports DeepSWE v1.1 in its place, as it did for the September build.
- MMLU-Pro: independent evaluation pending — No vendor value; Vals AI has not run this checkpoint (no model page on 2026-10-17).
- Terminal-Bench 2.x: independent evaluation pending — No Terminal-Bench 2.1 figure in the launch coverage, and no Vals AI run yet. The September build's 74.53 (Vals AI) is not carried over.
- MMMU-Pro: not reported by the vendor
API pricing
$0.15 per million input tokens, $0.6 per million output tokens (USD, cheapest OpenRouter endpoint, checked 2026-10-17)
Architecture and hardware
- Parameters
- 552B total, 16B active per token (MoE)
- Estimated VRAM at Q4
- ~318 GB, Frontier class
Links
More from DeepSeek
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.