| What | A tuned checkpoint of DeepSeek-V4.1-Flash, not a new model family |
|---|---|
| Released | October 17, 2026 (announced on X in the late afternoon Beijing time) |
| API name | deepseek-flash, unchanged. Legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work and now route to this checkpoint |
| Price | Unchanged: $0.15/1M input (off-peak, $0.30 peak), $0.60/1M output, $0.003/1M cached input |
| The claim | DeepSeek's words: a Pareto improvement in performance and cost. Better scores, faster completion, shorter outputs. |
| Architecture | Causal Encoder-Decoder sparse MoE: 552B-parameter backbone + 196B Engram, 8B parameters active on input / 16B on output |
| Context | 1M tokens in, up to 384K tokens out; native image understanding |
| Weights | MIT license, published on Hugging Face. Check the model card for this checkpoint's files. |
| Tech report | "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression" |
Cost calculator
The price did not move. Off-peak input is $0.15 per million tokens, peak is $0.30, output is $0.60, and cache reads are $0.003 per million. Cache reads matter here: agent loops resend long histories, and this model's compressed KV cache (890 bytes per token, roughly a quarter of the previous Flash generation) is built to make those resends cheap.
Estimates use DeepSeek's published rates. Peak/off-peak windows are defined in DeepSeek's docs; this calculator defaults to off-peak. One independent 56-trial pilot measured $0.0157 per finished coding task, versus $0.0577 for GPT-6.1 Sol and $0.0653 for Claude Sonnet 5.5. That is one pilot's result.
What changed
On October 17, 2026, DeepSeek posted that V4.1-Flash-1017 was live on its website, app, and API. It replaces the September 10 V4.1 Flash checkpoint under the same name and the same price. Three things moved:
- Scores. Every one of the 11 benchmarks DeepSeek reports went up, by 2.0 to 11.1 points. These are DeepSeek's numbers from test harnesses it chose; nobody has reproduced them independently yet. The biggest jump is Terminal-Bench 4.0, from 31.2 to 42.3.
- Speed and length. Tasks finish faster and outputs are shorter. DeepSeek gives no exact numbers, so treat this as a direction, not a percentage.
- Tuning. The checkpoint is tuned for DeepSeek's own agent harness (
dsh, v0.2.1-rc.1 or later recommended), Tencent's CodeBuddy, and WorkBuddy. Reinforcement learning with self-feedback targets end-to-end project work. DeepSeek's examples run from game creation to 3D modeling.
What did not change: the price, the API model name, the 1M-token context window, the MIT weight license, or the 384K-token max output.
Benchmarks
| Benchmark | V4.1-Flash-1017 | V4.1 Flash (Sept) | Claude Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| DeepSWE v1.1 | 77.0 | 74.2 | 74.0 | 73.0 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | n/a | n/a |
| HLE, no tools | 42.6 | n/a | 56.3 | n/a |
The 77.0 on DeepSWE v1.1 is the headline: ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) on that one coding benchmark, up from 74.2 on the September checkpoint. The counterweight is Humanity's Last Exam, a hard reasoning test, where 1017 scores 42.6 against Opus 5's 56.3. DeepSeek's own table also shows the new checkpoint beating Kimi K3 on 10 of 11 tests and GPT-5.6 Sol on 9, while trailing Opus 5 on 7.
On cost efficiency: an independent 56-trial pilot completed tasks at $0.0157 versus $0.0577 for GPT-6.1 Sol and $0.0653 for Claude Sonnet 5.5. Artificial Analysis measured the September checkpoint at 227 output tokens per second, about 3.4x faster than V4 Pro, and gave it a 39 on its Intelligence Index. Expect an updated index score for 1017 in the coming days.
How to access it
- Web and app. Live since October 17, 2026, free to use as before.
- API. Model name
deepseek-flash. Nothing else in your request changes:
curl https://api.deepseek.com/chat/completions \
-H "Authorization: Bearer $DEEPSEEK_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"deepseek-flash",
"messages":[{"role":"user","content":"Explain KV cache compression."}]}'- Sampling. DeepSeek recommends temperature 1.0 and top_p 1.0 for this checkpoint.
- Best with its own harness. DeepSeek says to use Harness v0.2.1-rc.1 or later "for the best experience with the new model." DeepSeek is building an end-to-end stack around its own harness.
- Weights. MIT-licensed on Hugging Face. The published weights are reported at 510 GB, so a local run is a server-class project. Check the model card before downloading; the 1017 files' availability was still being confirmed by trackers on day one.
Search & Fetch APIs
The second half of the announcement is infrastructure. DeepSeek opened its Search and Fetch APIs to all developers, running on an in-house search engine it says serves hundreds of millions of requests a day. The pitch is straightforward: a cheap coding model plus a first-party web layer behind one API key, aimed squarely at agent builders and at dedicated search-API providers.
- Price: $0.75 per 1,000 successful calls, on each endpoint.
- Free window: calls through server-side tools in DeepSeek Harness, Claude Code, and Codex stay free for DeepSeek API users until per-request billing starts on November 1, 2026, Beijing time.
- Limit: 5 QPS per account.
Use the calculator above to add search calls to your monthly estimate.
FAQ
What is DeepSeek V4.1-Flash-1017?
An updated post-training checkpoint of DeepSeek-V4.1-Flash, released October 17, 2026. It is served under the same deepseek-flash API name at the same price. DeepSeek says it scores higher, finishes tasks faster, and writes shorter outputs than the September checkpoint. The dated suffix follows DeepSeek's earlier pattern (V4-Flash-0731, V4-Pro-0813).
How much does it cost?
$0.15 per 1M input tokens off-peak ($0.30 at peak), $0.60 per 1M output tokens, and $0.003 per 1M cached input tokens. The price is unchanged from the September checkpoint. Flash input costs about one-fourth of DeepSeek V4 Pro and output about one-third.
Is it free to use?
The web app and mobile app are free, as before. The API is paid. If your volume is small, the free tiers of the web/app plus the calculator above will tell you whether the API is worth it.
How do I call it in the API?
Use the model name deepseek-flash. No code changes needed if you already used it. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work and are now routed to the new checkpoint. DeepSeek recommends temperature 1.0 and top_p 1.0.
Is it better than Claude Opus 5 or GPT-5.6 Sol?
On the coding benchmarks DeepSeek reports: yes. 77.0 on DeepSWE v1.1 versus 74.0 for Opus 5 and 73.0 for GPT-5.6 Sol. On hard reasoning it is not: 42.6 on Humanity's Last Exam against Opus 5's 56.3. And all of these are DeepSeek's own numbers, not independently reproduced.
Can I run it locally?
DeepSeek publishes the weights under an MIT license on Hugging Face. The reported size is 510 GB of weights, so this is a server-class project. The FP4 KV cache design (890 bytes per token) keeps the serving memory down, but the weights are the bill. Verify the 1017 checkpoint files on the model card before downloading.
What are the Search and Fetch APIs, and what do they cost?
First-party web search and page-fetch endpoints for agents, released the same day. Each costs $0.75 per 1,000 successful calls, capped at 5 QPS per account. Calls made through DeepSeek Harness, Claude Code, or Codex stay free for DeepSeek API users until November 1, 2026, Beijing time.
Sources
Benchmarks and release details are DeepSeek's own claims, cross-checked across independent write-ups from October 17–18, 2026:
- MarkTechPost: release details, full spec table, 1017-vs-September comparison
- AI Weekly: weights on Hugging Face, $0.75 Search API, 56-trial cost pilot
- Dealroom: "moves up the stack" framing, funding and IPO context
- TECHi: worked cost numbers ($0.15 off-peak, 0.9 GB cache per 1M tokens, 510 GB weights)
- The Straits Times: mainstream confirmation of the October 17 release
- Startup Fortune: the Harness-tuning angle
- OpenRouter: model directory entry: architecture, context, third-party pricing
- D-CSIL briefing: CED architecture, Harness v0.2.1-rc.1
Prices change. This page was last updated on October 18, 2026. Nothing here is affiliated with DeepSeek.
Machine-readable summary: llms.txt
Ask an AI about this guide: