DeepSeek Releases V4.1-Flash-1017, an Open-Weight Update Aimed at Coding Agents

On October 17, 2026, the Chinese AI lab DeepSeek posted V4.1-Flash-1017 to Hugging Face under an MIT license, an incremental update to the V4.1-Flash model it released in September. The update keeps the same architecture and API pricing but is tuned specifically for coding and agentic tasks, according to the model card DeepSeek published for the update on Hugging Face.

V4.1-Flash is a sparse mixture-of-experts model with a 552-billion-parameter backbone — about 763 billion parameters once its Engram memory module and vision encoder are counted — of which only 8 billion are active per token during input processing and 16 billion during output generation, per the Hugging Face model card. It supports a context window of up to 1 million tokens and uses a compression technique DeepSeek calls Compressed Sparse Attention 2, which the model card says cuts the key-value cache well below the size used by the prior Flash generation.

That cache reduction matters beyond a spec sheet: it lowers the GPU memory needed to hold long conversations or large codebases in context, which is what makes a 1-million-token window practical to run rather than theoretical.

What changed in the 1017 update

According to the comparison table on DeepSeek's own model card for the update, the 1017 revision raised the model's Terminal-Bench 4.0 score from 31.2 to 42.3, its ProgramBench score from 20.3 to 29.1, and its no-tools score on the Humanity's Last Exam reasoning test from 36.8 to 42.6, compared with the September version of V4.1-Flash. On the coding-agent benchmark DeepSWE v1.1, the updated card lists a score of 77.0, up from 74.2, and above the 74.0 that DeepSeek's own comparison table attributes to Anthropic's Opus 5 on the same test.

These are DeepSeek's own figures, generated on test setups the company chose. Tech City Authority, which reviewed the release, reported that the results are vendor-supplied and that no independent lab had reproduced them as of publication. The outlet also flagged that the model swap behind the deepseek-flash API alias happened automatically, without a version bump that developers would need to opt into — a detail that matters for anyone trying to reproduce an evaluation run before and after October 17, since the model answering a given API call changed without a corresponding change in the calling code.

Why it is aimed at coding agents specifically

DeepSeek is pairing the update with a Search and Fetch API, priced at $0.75 per 1,000 successful calls but free through October 31, 2026 (16:00 UTC) for traffic routed through DeepSeek's own agent harness, Anthropic's Claude Code or OpenAI's Codex, according to DeepSeek's release announcement. The model itself, the announcement says, was optimized for that same in-house harness together with Tencent's CodeBuddy and WorkBuddy, and DeepSeek recommends developers upgrade to DeepSeek Harness version 0.2.1-rc.1 or later to see the reported gains.

What it means for researchers and engineers

For students, researchers and engineers who cannot pay for closed frontier APIs, an MIT-licensed model with a 1-million-token context and the same per-token API pricing as before is directly usable: the weights can be downloaded, inspected and run locally through inference engines such as vLLM and SGLang rather than only through a hosted API. That openness also means anyone can attempt to reproduce the benchmark claims directly against the published weights, rather than taking DeepSeek's, or any outlet's, word for the scores.

The more immediate caution is for anyone building on the hosted API rather than the open weights: because the deepseek-flash endpoint now silently serves the 1017 checkpoint, evaluation pipelines, fine-tuned prompts or benchmark comparisons run against "DeepSeek V4.1-Flash" before and after October 17, 2026 are not necessarily comparing the same model, a distinction nothing on DeepSeek's release page or pricing page flags at a glance.