DeepSeek has shipped V4.1-Flash-1017, an update to its smallest current-generation model. It went out on October 17 under the existing API name, with no announcement event and no price change. The Hugging Face model card calls it a "minor update", but DeepSeek's own benchmark table shows large jumps on agentic coding and terminal tasks. The weights are again available under the permissive MIT license, which makes this one of the more notable open-weight releases of the month.
What DeepSeek Changed in Five Weeks
The original V4.1-Flash arrived on September 10 as the first and smallest member of a new DeepSeek architecture family. Barely five weeks later, the 1017 snapshot replaces it in DeepSeek's web and mobile apps and behind the deepseek-flash API endpoint. Developers don't need to change any code: according to DeepSeek's API change log, the calling method and pricing stay the same, and the older Flash model names now route to the new snapshot.
The model card says the update mainly reduces task execution time and output length. DeepSeek hasn't said by how much. That matters because shorter outputs lower the bill for each task even when the per-token price stays put. At the peak-hour rate of roughly $0.30 per million uncached input tokens and $1.20 per million output tokens, with off-peak rates half that, any cut in output volume goes straight into cost savings for heavy agent workloads.
DeepSeek says it tuned the release for agent software, including its own DeepSeek Harness. It also names two Tencent products: the CodeBuddy coding assistant and WorkBuddy, a desktop office agent. The accompanying technical report is titled "Pushing the Limits of KV Cache Compression". It points to a focus on making long-context inference cheaper. According to SandBase, the model card lists the model at 763 billion total parameters in a Mixture-of-Experts design.
The Benchmark Numbers
All of the scores below come from DeepSeek and have not been independently verified. Even so, the gap between versions is unusually wide for a point update:
- DeepSWE v1.1: 77.0% of tasks resolved, up from 74.2%
- Terminal-Bench 4.0: 42.3, up from 31.2
- ProgramBench: 29.1, up from 20.3
- Humanity's Last Exam (no tools): 42.6, up from 36.8
- Apex Bench v1.1: 72.1, up from 64.6
In DeepSeek's comparison table, the update beats Anthropic's Claude Opus 5 on DeepSWE (77.0 vs 74.0) and Apex Bench. Opus 5 still leads on seven of the eleven benchmarks DeepSeek published, though. That includes Terminal-Bench 4.0, ProgramBench and Humanity's Last Exam, where Claude's lead is close to 14 points. Note that Opus 5 is no longer Anthropic's newest model. SandBase also points out that the results depend on the evaluation harness: Terminal-Bench was run in the Claude Code framework, and DeepSWE used a separate scaffold.
New Search and Fetch APIs
DeepSeek also opened in-house Search and Fetch APIs alongside the model. Search returns up to 50 ranked web results per query. Fetch turns a single web page into Markdown, the format agents typically consume. Each costs $0.75 per 1,000 successful requests. SiliconANGLE reports that calls made through DeepSeek Harness, Claude Code or OpenAI's Codex are free until November 1. With these APIs, DeepSeek is offering more of the stack an agent needs, not just the model, at a time when web grounding is becoming a standard part of agent setups.
The update follows a rough week for the company. DeepSeek had an API outage of nearly five hours on October 15. It apologised, issued credits equal to the previous week's spending, and promised to remove single points of dependency. Some experts quoted by Bloomberg linked the outage to a shortage of computing capacity, a recurring issue for Chinese labs working under US export controls.
Why It Matters
For the open-weight model race, V4.1-Flash-1017 shows how quickly Chinese labs are iterating. A model this large, released under an MIT license, gets close to a recent frontier model on several agentic coding benchmarks, and anyone can download, fine-tune and self-host it. It arrives the same week Meta released Muse Spark Open, so open models competing on agent workloads now have strong entries from both the US and China.
The release also shows a shift in what labs are optimising for. The headline claims aren't about raw intelligence. They're about finishing tasks faster with fewer tokens, which is the figure that decides what it costs to run coding agents at scale. Enterprises comparing models should look closely at cost per completed task, not only per-token prices.
There are caveats. All the scores are self-reported, and they depend on the harness used. The update was also pushed into a live API endpoint without notice, so teams running production agents on deepseek-flash got different model behaviour overnight. That's useful when the change is an improvement, but it's a reminder to pin snapshots and re-run evaluations whenever a provider swaps models behind a stable name.
