Up 11.1 points on DeepSeek's own table, and anyone can use the weights commercially

DeepSeek released an update to its V4.1-Flash AI model on October 17, pairing commercially usable downloadable weights with a company-reported improvement on a test of practical computer tasks. ^[1][4]

The new version, V4.1-Flash-1017, is available through DeepSeek's website, app and API, the service developers use to connect software to its models. DeepSeek also announced availability through the model-sharing platforms Hugging Face and ModelScope. ^[4]

DeepSeek's benchmark table puts the new model's Terminal-Bench 4.0 score at 42.3, against 31.2 for the previous version. That is an increase of 11.1 points, or about 36% relative to the earlier score, according to the company's figures. ^[1]

For developers choosing a coding tool, the release combines two attractions: a reported gain in completing work and permission to run or adapt the model themselves. The weights, the numerical settings learned during training, are available under the permissive MIT License. ^[1][3]

DeepSeek describes the release as a "minor update" and says it completes tasks faster while producing shorter outputs. ^[1][4] The practical question is whether those gains make it a better tool for finishing a developer's work, rather than merely a better entry in a benchmark table.

A test of finished work

Terminal-Bench tests an AI agent, a model connected to tools that let it take actions, inside an isolated command-line environment. The agent must carry out a task through the terminal, where users type commands to run programs and manage files. Credit depends on the finished result passing a check. ^[2]

Its 66 tasks are contributed by the community and reviewed by maintainers. They cover software, science, machine learning, operations, hardware, security and media. That makes the test broader than coding alone, even though it is useful for judging models offered as programming assistants. ^[2]

The attraction is concrete: the test asks whether an agent can finish realistic, multi-step work. A plausible explanation is not enough. The final result has to satisfy the task's verifier, the check used to decide whether the work succeeded. ^[2]

The setup around the model matters, too. A testing harness is the software that connects a model to its tools and manages its attempts to complete a task. DeepSeek says it used Claude Code, Anthropic's coding tool wired into the test, as its harness. Vals AI, the independent evaluator publishing a Terminal-Bench 4.0 leaderboard, uses mini-swe-agent, a minimal open-source tool that connects a model to a terminal. ^[1][2]

Vals AI's method averages results across three runs, with one attempt at each task in every run. Its published leaderboard does not establish an independent score for V4.1-Flash-1017, and DeepSeek's result should not be treated as directly comparable with scores produced using Vals AI's different setup. The reported speed and output-length improvements also await independent testing. ^[1][2][4]

The 11.1-point increase is therefore a substantial improvement in DeepSeek's own table, rather than an independently established position against rival models. ^[1][2] A developer evaluating the release should test the complete system they intend to use, because that is what will be doing the work.

Permission to build, with conditions

The licensing offer is straightforward. DeepSeek says both the repository and the model weights carry the MIT License. That gives developers an alternative to accessing the model solely through the company's hosted service: they can download the weights and run them with compatible software. ^[1]

The licence permits use, copying, modification, distribution, sublicensing and sale, including commercial use. Copies or substantial portions must retain the copyright and permission notices. The material comes without a warranty, and the licence limits liability. ^[3]

For a business, those permissions leave room to adapt the model, incorporate it into a product or distribute a modified version without a non-commercial restriction in the licence. ^[3] This is the release's most durable offer. A benchmark score describes a result under particular test conditions. The licence sets out what a developer is allowed to do with the released material.

Permission does not supply the computing infrastructure. DeepSeek's September 10 announcement described the preceding V4.1-Flash as a 552-billion-parameter model, with eight billion parameters active for input and 16 billion for output. Parameters are the learned numerical settings, and the active counts describe the portion used during those stages. ^[5]

That scale makes deployment a separate decision from downloading. Developers weighing a self-run installation need to consider compatible software and computing capacity alongside the freedom the licence provides.

Shorter answers, a separate cost question

DeepSeek's model card includes instructions for using the weights with Transformers, vLLM and SGLang, software tools used to load and run models. It also supplies a reference Python implementation and a Rust toolkit with Python bindings for preparing prompts and interpreting outputs. The release does not include a Jinja-format chat template, a commonly used way of defining how conversations are formatted for a model. ^[1]

DeepSeek characterises the update as an improvement in performance and cost, alongside its claims of faster task completion and shorter outputs. Its release announcement does not provide a before-and-after API price comparison, so it does not establish that prices are unchanged. ^[4]

Shorter output and lower prices are separate propositions. For developers, the useful measure is the total cost of getting an acceptable task finished, including any retries or corrections, rather than the length of a single answer.

The next test for V4.1-Flash-1017 is reproducibility: whether an independent evaluator can confirm its performance under the same harness and conditions used for other models. Until then, developers have weights they can use commercially and a company benchmark result they can put to the test. ^[1][2][3]