DeepSeek V4.1 Flash 1017: 77% on DeepSWE, MIT Weights, Same API Price

DeepSeek has shipped DeepSeek-V4.1-Flash-1017, a post-training update to its V4.1 Flash model. It is live on DeepSeek’s website, app, and API now, and the weights are on Hugging Face and ModelScope under the MIT license. DeepSeek reports 77.0% on DeepSWE v1.1 and a jump from 31.2 to 42.3 on Terminal-Bench 4.0, with shorter outputs and faster task completion than the version it replaces. API prices did not change.

The release went straight to the top of Hacker News, where the model card and the OpenRouter listing both cleared 500 points and the main thread passed 600 comments. Every benchmark number below is DeepSeek’s own. Nobody outside the company has reproduced them yet, and 9to6AI has not hands-on tested the model.

Deployability and license first

• License: MIT for both the repository and the weights. Commercial use, fine-tuning, and redistribution are allowed with the usual MIT notice.
• Hosted: On the DeepSeek API, the model name deepseek-flash now serves V4.1-Flash-1017. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp also route to it. If you already call DeepSeek, you have been moved to the new model without opting in.
• Third-party: OpenRouter lists it as deepseek/deepseek-v4.1-flash-1017 with a 1M-token context.
• Self-host: The checkpoint is roughly 510 GB across 48 safetensors shards, stored in FP8 with FP4 experts. DeepSeek’s reference inference code converts weights for 8-way tensor parallelism and runs with torchrun. This is a multi-GPU server model, not a laptop model.
• Same architecture: The config.json is identical to the original V4.1 Flash. DeepSeek describes that model as a 552B-parameter mixture-of-experts with about 8B active parameters for input and 16B for output (Hugging Face’s raw tensor count reads about 763B; we use DeepSeek’s figure). Context is 1M tokens, and vision input is native.

What changed

DeepSeek calls 1017 a “minor update” with “substantial improvements,” a phrase HN readers enjoyed. The company credits reinforcement learning with self-feedback for better end-to-end project work, and shows demos that include a 3D model of the Forbidden City, a browser racing game, and a Portal 2 recreation built in the Godot engine.

The other claim is efficiency: fewer output tokens and less wall-clock time per task, which DeepSeek frames as a Pareto improvement on performance and cost. It did not publish token or latency numbers to back that up, so treat it as a claim to test.

The benchmark table (DeepSeek-reported)

Terminal-Bench 4.0 ran in the Claude Code harness and DeepSWE in the Mini-SWE-V2 harness, according to DeepSeek’s footnotes.

How to read it:

• The jump over the old Flash is real on paper. The biggest gains are on Terminal-Bench (+11.1), ProgramBench (+8.8), and Apex Bench (+7.5). For a same-architecture update, that is a lot.
• Claude Opus 5 still leads on the hardest rows. Terminal-Bench, ProgramBench, NL2Repo, and HLE all favor Opus 5, some by wide margins.
• The comparison set is not the newest. The table uses GPT-5.6 Sol and Claude Opus 5, not OpenAI’s GPT-6.1 Sol or Anthropic’s Claude Opus 5.5. Expect the gap to the true frontier to be wider than this table suggests.

The cost chart is the real pitch

DeepSeek’s second chart plots DeepSWE v1.1 pass rate against average cost per task, using public leaderboard data for the other models. Read from that chart, V4.1-Flash-1017 at max effort reaches 77% for around $0.20 per task, and its low setting lands in the low 70s for close to $0.10. GPT-6.1 Sol peaks in the mid-70s at several times that cost, and Claude Opus 5 sits in the low-to-mid 70s at more than $10 per task.

That is the story for builders: not “best model,” but near-frontier coding results at a price that changes how many agent runs you can afford. The chart is DeepSeek’s, so verify the cost math on your own workloads.

API pricing and the new Search and Fetch APIs

Prices match what V4.1 Flash already charged. Per 1M tokens:

Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. Everything else, including full weekends, bills at the off-peak half price. Max output is 384K tokens, the concurrency limit is 2,500, and the API speaks OpenAI Chat Completions, the Responses API, and an Anthropic-format endpoint, with tool calls, JSON output, and vision.

DeepSeek also opened its Search and Fetch APIs to everyone. Search returns structured results from DeepSeek’s in-house engine, and Fetch returns a single page as Markdown. Each costs $0.75 per 1,000 successful requests. Server-side search and fetch calls made through DeepSeek Harness, Claude Code, or Codex on a DeepSeek API key are free for a limited window before per-request billing starts; the cutover is listed on DeepSeek’s pricing page.

Thinking mode is on by default at “high” effort, with “low” and “max” as the other API settings. The open-weight prompt format goes further and accepts a numeric reasoning budget from 1 to 100.

What builders should know

1. You may already be on it. Anyone calling deepseek-flash or the legacy V4 Flash names got switched. Rerun your evals, and check anything that depends on output length or formatting, since DeepSeek says responses are shorter now.
2. Self-hosting needs new prompt plumbing. There is no Jinja chat template in this release. Use DeepSeek’s encoding.py reference or its new deepseek-recipe Rust libraries (with Python bindings), which convert Chat Completions, Messages, and Responses requests into the V4.1 prompt format. V4.1 also changed the tool-call tag format compared with V4, so older parsers can break.
3. Schedule batch work off-peak. Evals, bulk refactors, and nightly agent jobs cost half outside the peak windows.
4. Check your data policy before you route production traffic. The hosted API is run by a China-based company. If that is a problem for you or your clients, the MIT weights let you self-host or use a third-party host instead.
5. Pair it, don’t just swap it. A common pattern in the HN thread was using DeepSeek Flash as the cheap workhorse and a frontier model as reviewer. One commenter’s code-review test found GPT-6.1 Sol faster and catching more issues, while another reported the max-effort setting taking over 11 minutes on a single drawing prompt. Cheap per token does not always mean fast per task.

We covered the wider open-weight race when Reflection released Beam, whose own table placed DeepSeek V4.1 Flash ahead on several tests: Reflection Beam: a 501B open-weight model built for efficiency. This update widens that lead on DeepSeek’s numbers.

What to watch

• Independent results on Terminal-Bench 4.0 and DeepSWE from leaderboard maintainers, not just DeepSeek.
• Whether the “shorter outputs” claim holds up in real agent traces, since that drives most of the cost story.
• Support in vLLM, SGLang, and other serving engines for this checkpoint and the deepseek-recipe format.
• Whether DeepSeek ships V4.1 Pro, which it has referenced but not yet launched.

Key takeaways

• DeepSeek-V4.1-Flash-1017 is a post-training update on the same 552B MoE architecture, with MIT-licensed weights.
• DeepSeek reports 77.0% on DeepSWE v1.1 and 42.3 on Terminal-Bench 4.0, up from 74.2 and 31.2. Claude Opus 5 still leads on the hardest benchmarks.
• API pricing is unchanged: $0.15 in / $0.60 out per 1M tokens off-peak, double at peak.
• deepseek-flash now points to the new model, so existing users should re-test.
• New Search and Fetch APIs cost $0.75 per 1,000 calls each.
• All numbers are vendor-reported, and 9to6AI has not tested the model yet.

Sources

• DeepSeek API news: DeepSeek-V4.1-Flash-1017 release
• Hugging Face model card: deepseek-ai/DeepSeek-V4.1-Flash-1017
• DeepSeek API: Models and pricing
• DeepSeek API: Search and Fetch guide
• DeepSeek API news: original DeepSeek-V4.1-Flash release
• OpenRouter: DeepSeek V4.1 Flash 1017
• Hacker News discussion