A dated Flash cut, not a new V4 paper. DeepSeek's 1017 tag is the buyer-facing weight bump after the September KV-cache Record.
DeepSeek posts V4.1-Flash-1017 on Hugging Face
DeepSeek-AI published deepseek-ai/DeepSeek-V4.1-Flash-1017 on Hugging Face around 17 October 2026. The card calls it a minor update to DeepSeek-V4.1-Flash with improvements across reasoning, coding, agentic tasks, and visual benchmarks, plus shorter outputs and faster task completion on the company's framing. Model size on the card: 763B parameters. Context window guidance: 1M tokens. License: MIT for the repository and weights.
Company table on the card (selected rows): Terminal-Bench 4.0 at 42.3 for 1017 versus 31.2 for prior Flash (Claude Code harness noted); DeepSWE v1.1 at 77.0 versus 74.2 (Mini-SWE-V2 harness); ProgramBench 29.1 versus 20.3; Automation-Bench v1.0.6 at 60.5 versus 54.8; Agents' Last Exam 35.4 versus 31.8. Comparator columns include K3, GLM-5.3, GPT-5.6 Sol, and Opus 5 with DeepSeek's numbers. Recommended sampling: temperature 1.0, top_p 1.0, max_tokens at least 256K.
The card says the release does not ship a Jinja chat template; an encoding folder holds a Python reference, and deepseek-recipe Rust/Python bindings are pointed for production prompt encoding. An inference folder covers weight conversion and local run notes.
Who the MIT tag binds
DeepSeek-AI shipped the weights and the company bench table. Self-host operators who pull 1017 accept MIT terms and their own serve stack. Teams still on the September V4.1-Flash KV-cache report are not automatically on 1017 until they change the revision. Closed API buyers only move if DeepSeek's product surfaces point at this tag.
A Flash revision with a printed agent delta
A public MIT 1017 checkpoint with company-claimed agent and coding lifts versus the prior Flash row, plus encoding/recipe pointers for non-Jinja prompts. It is not a new DeepSeek-V4 architecture paper. It is not third-party confirmation of Terminal-Bench 42.3.
Harness and cost are DeepSeek's
Date: Hugging Face release window 17 October 2026. License: MIT. Price: self-host is your GPUs; API list rates are not on this card. Region: export and deemed-export review remains buyer-side. Independent reproduction of Terminal-Bench 4.0 at 42.3 or DeepSWE at 77.0 under the same harnesses is not in the card. Relation to the September arXiv KV-cache report is lineage, not a fresh FLOPs plot on this page.
Pin 1017 and re-bench before you cut token budget
Platform owners should pin the 1017 revision, load the encoding path DeepSeek documents, and re-run their own agent holdouts before assuming shorter outputs cut billable tokens. Keep the September KV-cache Record for architecture; use this Dispatch for the weight tag. Ignore "Pareto frontier" as a contract term until your harness matches theirs.