DeepSeek-V4.1-Flash-1017 (also searched as deepseek flash 1017 / deepseek v4.1 flash 1017) is a minor update to DeepSeek-V4.1-Flash. DeepSeek’s Hugging Face model card and API changelog describe stronger reasoning, coding, agentic, and visual results than the prior Flash checkpoint, with shorter outputs and faster task completion—framed as a Pareto step on performance versus cost. Open weights are on Hugging Face and ModelScope; the hosted API continues under deepseek-flash at unchanged calling method and pricing.
Quick facts
- Weights: deepseek-ai/DeepSeek-V4.1-Flash-1017 (also ModelScope under the same org/name)
- License (HF repo & weights): MIT, per the model card
- API model id:
deepseek-flash— DeepSeek’s Oct 17, 2026 changelog states the API calling method and pricing remain unchanged - Base Flash product news (Sept 10, 2026): smallest model in the new architecture family, with native visual understanding; marketed as a Causal Encoder–Decoder MoE with large KV-cache compression vs prior generation
What Flash-1017 changes vs prior Flash
Per the Hugging Face introduction for Flash-1017, the update delivers substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks, while reducing task execution time and output length so users can finish work faster and at lower cost. DeepSeek’s Oct 17 API changelog similarly calls out a performance–cost Pareto improvement and says the official web and APP models were updated to the same checkpoint.
Tooling notes from DeepSeek:
- Optimized for DeepSeek Harness, Tencent CodeBuddy, WorkBuddy, and similar platforms
- Upgrade DeepSeek Harness to v0.2.1-rc.1 or later for the best experience with the new model
- API news also highlights end-to-end project demos (games / 3D scenes) trained with reinforcement learning and self-feedback
Separately on Oct 17, DeepSeek opened its in-house Search & Fetch APIs to everyone. The Flash-1017 news post lists $0.75 per 1,000 calls per endpoint. Search & Fetch via server-side tools in DeepSeek Harness, Claude Code, and Codex are described as free for DeepSeek API users until billing starts Oct 31, 2026, 16:00 UTC.
Vendor benchmarks (label as such)
The following numbers are DeepSeek’s published comparisons on the Flash-1017 Hugging Face card and mirrored in the API changelog—not independent third-party audits. Use them as directional product data and re-run on your harness.
| Benchmark | DS-V4.1-Flash-1017 | DS-V4.1-Flash (prior) |
|---|---|---|
| HLE w/o tools | 42.6 (44.0†) | 36.8 (39.1†) |
| Terminal-Bench 4.0 | 42.3 | 31.2 |
| DeepSWE v1.1 | 77.0 | 74.2 |
| ProgramBench | 29.1 | 20.3 |
| NL2Repo-Bench | 68.2 | 65.4 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 |
| Agents’ Last Exam | 35.4 | 31.8 |
| Apex Bench v1.1 | 72.1 | 64.6 |
| Chartography (w/tools) | 83.0 | 78.9 |
| BabyVision (w/tools) | 92.5 | 89.6 |
| ZeroBench-main (w/tools) | 51.0 | 49.0 |
† Text-only subset of HLE. DeepSeek notes Claude Code harness for Terminal-Bench 4.0 and Mini-SWE-V2 for DeepSWE v1.1; ZeroBench-main is listed as Pass@5 in the changelog. The HF card also lists competitor columns (K3, GLM-5.3, GPT-5.6 Sol, Opus 5)—treat those as vendor-collated figures.
API still deepseek-flash
When DeepSeek launched V4.1-Flash (Sept 10, 2026), developers were told to set the model to deepseek-flash. Legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp were retired but temporarily routed to V4.1-Flash. The Oct 17 changelog is explicit: after the Flash-1017 upgrade, “The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model.”
Practical takeaway: most apps keep the same model string and receive the new checkpoint on DeepSeek’s hosted endpoints. If you need weight-level reproducibility or offline inference, download the open HF/ModelScope checkpoint rather than expecting an older Flash snapshot id on the public API.
Local / open-weight notes
The HF card says this release does not include a Jinja chat template; an encoding/ folder provides a Python reference encoder for multi-turn chat, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages, and interleaved images. DeepSeek also points to deepseek-recipe (Rust libraries with Python bindings) for production prompt encoding across Messages / Chat Completions / Responses-style requests.
Recommended sampling parameters listed on the card:
| Parameter | Value |
|---|---|
temperature |
1.0 |
top_p |
1.0 |
context_window |
1M tokens |
max_tokens |
≥ 256K |
See the repo’s inference folder for weight conversion and local run instructions. Size reporting differs by source: the September product news described a 552B-parameter MoE with asymmetric active params (8B active for input / 16B for output in DeepSeek’s wording); the Hugging Face safetensors UI may show a different aggregate count—cite the page you mean when quoting parameters.
Cost context
Flash-1017 itself does not introduce a new public price list in the Oct 17 note (pricing “remain unchanged”). The September V4.1-Flash announcement said Flash’s KV cache needs about 1/4 the HBM and 1/8 the SSD storage versus the previous generation, enabling lower API prices with peak/off-peak pricing (off-peak at 50% of peak). Confirm live token rates on DeepSeek’s Models & Pricing docs before budgeting; treat shorter outputs on Flash-1017 as a potential cost lever only after you measure your own workloads.
Bottom line
Flash-1017 is DeepSeek’s mid-cycle Flash refresh: open weights on HF/ModelScope, vendor benchmarks up versus prior Flash (notably Terminal-Bench 4.0 and several agent suites), and hosted access still via deepseek-flash with unchanged API pricing language. Label DeepSeek’s score tables as vendor results, keep Harness updated, and watch Search & Fetch billing after Oct 31, 2026 UTC.
Sources
- Hugging Face: DeepSeek-V4.1-Flash-1017
- DeepSeek API Docs: Change Log (2026-10-17 Flash-1017)
- DeepSeek API Docs news: DeepSeek-V4.1-Flash-1017
- DeepSeek: Introducing DeepSeek-V4.1-Flash
Sources verified Oct 17, 2026 (Pacific Time)
