DeepSeek Ships V4.1-Flash-1017 Under MIT -- a 'Minor Update' That Lifts Its Own Terminal-Bench Score From 31.2 to 42.3 and Opens Its Search Engine as a Paid API

DeepSeek's release note for October 17, 2026 is short: "Stronger benchmark performance, faster task completion, and shorter outputs than the previous version -- a Pareto improvement in performance and cost." The Hugging Face model card calls DeepSeek-V4.1-Flash-1017 "a minor update to DeepSeek-V4.1-Flash", which shipped on September 10, and says it "reduces task execution time and output length". The repository and weights are under the MIT License. The Hacker News thread for the release passed 700 points and 590 comments within a day.

What changed for API users

• Same name, new model. DeepSeek's API docs say deepseek-flash "is now served by DeepSeek-V4.1-Flash-1017, and the calling method remains unchanged." The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but "the corresponding models have been retired", and those calls are billed at the Flash price.
• Search becomes a product. DeepSeek opened its in-house search engine as two endpoints, Search and Fetch: "Each endpoint: $0.75 per 1,000 calls." Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free for now; "Billing starts Oct 31, 2026, at 16:00 UTC."
• A harness of its own. The note says the model is "Optimized for DeepSeek Harness, Tencent CodeBuddy, WorkBuddy, and more", and asks users to upgrade to DeepSeek Harness v0.2.1-rc.1 or later. The harness is in developer preview.
• No chat template. The card says "This release does not include a Jinja-format chat template." DeepSeek instead ships a Python reference encoder and deepseek-recipe, Rust libraries with Python bindings that convert Messages, Chat Completions and Responses API requests into its prompt format. Recommended sampling is temperature 1.0 and top_p 1.0, with max_tokens of at least 256K and a 1M-token context.

What the benchmark table claims

DeepSeek's table compares 1017 with the September Flash release, Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. Its largest jump is on Terminal-Bench 4.0, from 31.2 to 42.3, run in the Claude Code harness. That is ahead of GLM-5.3 (37.9) and GPT-5.6 Sol (39.9), but behind Opus 5 (51.8). On DeepSWE v1.1 it reports 77.0 against Opus 5's 74.0. On Automation-Bench, Agents' Last Exam and Apex Bench it posts the highest number in the table: 60.5, 35.4 and 72.1. It trails Opus 5 on ProgramBench (29.1 vs 37.0) and NL2Repo-Bench (68.2 vs 75.3). These are DeepSeek's own runs, and several of the harnesses and benchmark versions are new enough that few independent scores exist yet.

The release note's demos lean on long, open-ended builds. DeepSeek says the model "uses reinforcement learning with self-feedback to improve end-to-end project development", and shows a 3D Forbidden City, a web racing game and a Portal 2 recreation in Godot.

What remains uncertain. DeepSeek does not quantify "shorter outputs" or "faster task completion" in the release note, so the cost claim cannot yet be checked per task. The card publishes no new parameter count for this checkpoint. Teams pinned to deepseek-flash got a different model on October 17 without changing a line of code, and the API docs list no alias that keeps the September model.

Analysis: The more consequential change may be the search API, not the weights. A lab that sells a cheap model, a harness, and its own web search and fetch is assembling the full agent stack that Anthropic, OpenAI and Google sell, at Flash prices. Free server-side search inside Claude Code and Codex until October 31 is a direct bid for agent traffic that today goes to other providers' tools. For anyone running evaluations, the silent model swap behind a stable name is a reminder to log the served model version with every result, because "deepseek-flash" on October 16 and October 18 are different systems.