Software tools that help people write and run code keep changing in small steps.
While many Western labs continue to ship new models at a fast pace, DeepSeek has taken a different path with its latest release.
Instead of launching a wholly new model, the company updated one of its existing ones to 'DeepSeek-V4.1-Flash-1017.'
The company describes it as a minor update to the version it released in September. The new checkpoint is available on the DeepSeek website, app, and API under the same model name used before. Weights are also posted under an open license on Hugging Face and ModelScope.
The model keeps the same overall design as the earlier Flash release.
It is a mixture-of-experts system built around a causal encoder-decoder layout.
The backbone holds about 552 billion parameters, and a separate conditional memory component adds roughly 196 billion more. Only a small share of those parameters activate on any given token: about 8 billion while the model reads input and about 16 billion while it writes output. It accepts both text and images and can produce up to a million tokens of context.
Recommended sampling settings on the model card are a temperature of 1.0, a top-p of 1.0, and a maximum output length of at least 256 thousand tokens.
Read: DeepSeek, The Chinese AI That Makes Silicon Valley Nervous And The U.S. Concerned
DeepSeek’s own benchmark table shows higher scores than the September checkpoint across the eleven tests it lists.
On Terminal-Bench 4.0 the score rose from 31.2 to 42.3. On DeepSWE v1.1 it moved from 74.2 to 77.0. Other coding and agent tests such as ProgramBench, NL2Repo-Bench, Automation-Bench, and Apex Bench also increased.
Visual tests that involve charts or simple image tasks showed smaller gains.
The table places the new checkpoint ahead of several other open or closed models on some agent and coding measures, while it remains behind Claude Opus 5 on raw reasoning tasks such as Humanity's Last Exam without tools and on a few of the harder coding suites. All of the numbers come from DeepSeek’s evaluations, run with specific harnesses it chose.
The company also states that the update finishes tasks faster and produces shorter outputs than the prior version.
It attributes part of the improvement in end-to-end project work to reinforcement learning that includes self-feedback.
Public examples shown with the release include a 3D model of the Forbidden City, a simple web racing game, a recreation of a puzzle game built in the Godot engine, and a multi-era landscape scene.
Pricing on the DeepSeek API did not change with this release. Off-peak rates remain $0.15 per million input tokens when the cache misses, $0.003 when the cache hits, and $0.60 per million output tokens.
Peak rates are twice those figures during two weekday windows in UTC.
At the same time DeepSeek opened Search and Fetch endpoints to general API users.
Each endpoint is listed at $0.75 per thousand calls. Calls made through server-side tools inside DeepSeek Harness, Claude Code, or Codex stay free for DeepSeek API users until 31 October 2026 at 16:00 UTC.
The model card notes that the release does not ship a standard chat template; instead it provides a Python reference script and supporting libraries for encoding multi-turn conversations, tool calls, and image content.
Inference code for weight conversion is included in the repository. Users who already call the deepseek-flash endpoint receive the new checkpoint without changing their code.
Independent checks of the reported scores had not appeared in the first day or two after the release.
Some early side-by-side tests by third parties compared the model against earlier DeepSeek versions on small coding tasks and noted shorter run times and lower token use, consistent with the company's description of reduced output length.
The weights themselves are large; the public checkpoint is on the order of half a terabyte, so local use remains limited to multi-GPU machines or specialized inference setups. For most people the practical path is the hosted API or the web and app interfaces.
The update arrives while many groups continue to measure how well these models handle multi-step coding work, terminal commands, and tool use over long contexts.
The numbers DeepSeek published, and the fact that the weights are open, give other researchers a concrete checkpoint they can examine or try to reproduce. Whether the measured gains hold under different harnesses or on tasks outside the published suite is something later tests will show.
DeepSeek has also been building out the software layer that sits around its models. DeepSeek Harness is an open-source agent runtime that connects a model to files, a terminal, tools, and session history so it can keep working across many steps.
The project follows an everything-is-a-plugin design, meaning the model adapter, tools, session store, sandbox, and even the agent loop itself can be swapped or extended. A desktop app for macOS and Windows reached a public preview in late September, and the company has said the new Flash checkpoint is optimized for it. Users of the official DeepSeek API have been directed to update to a recent version of Harness for the best results with the October release.
At the same time, DeepSeek has faced growing questions about how quickly it can keep pace.
Service outages have become more frequent this year, including a multi-hour disruption in mid-October that affected both the chat product and the API.
Some observers note that the company has shifted from the rapid model drops that marked its earlier years toward smaller updates and infrastructure work. The gap with the fastest-moving Western labs on raw capability has narrowed on some measures, yet the pace of new frontier releases from those labs has not slowed.
Still, DeepSeek continues to ship open weights, low prices, and agent tooling that other groups can inspect and build on.
Read: China Races To Put DeepSeek Into Everything: How The U.S. Failed To See The First AI War Coming
