A model that activates 16 billion parameters per token solves 77.0% of DeepSWE v1.1, three points ahead of Claude Opus 5, for under $0.20 per task according to DeepSeek's own cost chart. It's DeepSeek-V4.1-Flash-1017, released on October 17, 2026 on DeepSeek's website, app and API, with the weights on Hugging Face under the MIT license.
This isn't a new model but a minor update to V4.1-Flash. The gains come from post-training: better coding and agent scores with shorter answers, which makes every task cheaper.
What the 1017 release brings
The architecture is V4.1-Flash's: a multimodal (text and image) MoE with a 552-billion-parameter backbone and a one-million-token context window. What sets it apart is its Causal Encoder-Decoder design: 20 encoder layers and 20 decoder layers, with the decoder's global KV cache projected from the encoder. That way it activates 8 billion parameters per token while reading input and 16 billion while generating, and keeps the KV cache at 890 bytes per token, a quarter of V4-Flash. Agent workloads read far more than they write, so that asymmetry is exactly what drives the cost per task down.
It also keeps the adjustable reasoning effort, set as an integer from 1 to 100 rather than the usual three or four levels.
How it stacks up against frontier models
Figures from the model card, at maximum reasoning effort:
| Benchmark | V4.1-Flash-1017 | V4.1-Flash | GPT-5.6 Sol | Claude Opus 5 |
|---|---|---|---|---|
| DeepSWE v1.1 | 77.0 | 74.2 | 73.0 | 74.0 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 39.9 | 51.8 |
| NL2Repo-Bench | 68.2 | 65.4 | 56.8 | 75.3 |
| Automation-Bench | 60.5 | 54.8 | 45.8 | 50.3 |
| HLE without tools | 42.6 | 36.8 | 44.5 | 56.3 |
Read honestly, it wins at fixing issues and at automation, and still trails Opus 5 by a wide margin on long terminal tasks, building a repository from scratch and pure reasoning. Where the gap is huge is price: on DeepSeek's DeepSWE chart the model hits 77% for just under $0.20 per task, and in low-effort mode it clears 72% for about $0.10. Its predecessor cost around $0.50, and the rivals on the chart land between $2 and more than $10 per task.
What it means if you build with agents
If you run a coding agent on well-scoped tickets, a model that gets close to frontier quality at a tenth of the cost changes how many parallel attempts you can afford. On the previous version's card, DeepSeek also reports results across harnesses: 69.8% on DeepSWE with Claude Code versus 74.2% with mini-SWE, so the tool wrapped around the model matters as much as the model itself.
Alongside the model, DeepSeek is opening the web Search and Fetch APIs it uses internally to all developers, at $0.75 per 1,000 calls, and free as server-side tools in DeepSeek Harness, Claude Code and Codex until October 31, 2026.
For self-hosting, the MIT license is the good news and the size is the bad: the weights add up to 763 billion parameters in FP8 and 8-bit, and the repository ships no Jinja chat template. The prompt format lives in a Python reference implementation and in deepseek-recipe, a set of Rust libraries.
The caveat that matters here: the table compares against Opus 5 and GPT-5.6 Sol, not Claude Opus 5.5, Anthropic's latest model, and every number comes from DeepSeek. Until independent evaluations land, the only test that counts is running it against your own tickets.
Sources
- DeepSeek-V4.1-Flash-1017 (model card), DeepSeek on Hugging Face (2026-10-17)
- DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression, DeepSeek on Hugging Face
- Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs, The Decoder (2026-10-17)
- DeepSeek-V4.1-Flash-1017 (discussion), Hacker News (2026-10-17)