DeepSeek Releases DeepSeek-V4.1-Flash-1017 Open-Weight Update

DeepSeek has published DeepSeek-V4.1-Flash-1017 on Hugging Face — a minor update to DeepSeek-V4.1-Flash, released under the MIT licence. The model card was created on 17 October 2026 (02:40 UTC per Hugging Face’s API). DeepSeek does not give a parameter count on the card; none is stated here.

All benchmark numbers below are from DeepSeek’s own comparison table on that card. They are not independently verified.

What DeepSeek claims

DeepSeek writes that V4.1-Flash-1017 is “a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks.” It also says the model “reduces task execution time and output length, enabling users to complete tasks faster and at lower cost.” No milliseconds or percentage speed-ups are published — only that qualitative claim.

Recommended sampling on the card: temperature 1.0, top_p 1.0, context window 1M tokens, max_tokens ≥ 256K. There is no Jinja chat template; DeepSeek points to a Python reference in encoding/encoding.py and to deepseek-recipe (Rust libraries with Python bindings) for production encoding of Messages / Chat Completions / Responses-style requests.

DeepSeek’s table vs the prior Flash

Selected rows from DeepSeek’s card (1017 vs V4.1-Flash):

DeepSeek notes it used the Claude Code harness for Terminal-Bench 4.0 and Mini-SWE-V2 for DeepSWE v1.1. The same table also lists K3, GLM-5.3, GPT-5.6 Sol and Opus 5.

Where Opus 5 is still ahead (DeepSeek’s numbers)

On DeepSeek’s own table, Claude Opus 5 scores higher than V4.1-Flash-1017 on several rows, including:

• HLE w/o tools: Opus 5 56.3 vs 1017 42.6 (text-only subset: Opus not listed separately; 1017’s † figure is 44.0)
• Terminal-Bench 4.0: Opus 5 51.8 vs 1017 42.3
• ProgramBench: Opus 5 37.0 vs 1017 29.1
• NL2Repo-Bench: Opus 5 75.3 vs 1017 68.2
• Chartography (w/tools): Opus 5 84.0 vs 1017 83.0
• BabyVision (w/tools): Opus 5 94.1 vs 1017 92.5
• ZeroBench-main (w/tools): Opus 5 52.0 vs 1017 51.0

1017 does lead Opus 5 on other rows in the same table — for example DeepSWE v1.1 (77.0 vs 74.0), Automation-Bench (60.5 vs 50.3), Agents’ Last Exam (35.4 vs 28.6) and Apex Bench v1.1 (72.1 vs 65.8). That is still DeepSeek scoring DeepSeek. Do not read the release as an overall win over Opus 5.

Why it matters

Open Chinese Flash-class weights remain a moving target for agent harnesses. A “minor” dated drop that lifts Terminal-Bench and DeepSWE on the vendor’s sheet will be picked up quickly in eval stacks — including work such as OpenGCUA’s multi-model computer-use study, which already names DeepSeek-V4.1-Flash. The useful discipline is to keep Opus-ahead rows in view whenever the card is quoted.

For prior DeepSeek context, see TSN on DeepSeek V4 and China’s vertical AI stack and reported funding colour.

What’s still unknown

• Independent scores under the same harnesses
• Any official parameter count from DeepSeek (none on the card)
• How much of the “faster / shorter” claim shows up outside DeepSeek’s own tooling

What this means

• DeepSeek-V4.1-Flash-1017: MIT open weights; minor update to V4.1-Flash; HF created 17 Oct 2026.
• DeepSeek’s table: broad gains vs prior Flash on agent and vision benches; Opus 5 still ahead on HLE, Terminal-Bench 4.0, ProgramBench, NL2Repo-Bench and several vision rows.
• Serving notes: 1M context, max_tokens ≥ 256K; encoding via encoding.py / deepseek-recipe, not Jinja.
• Benchmarks are DeepSeek’s own — not independent.

Sources

• deepseek-ai/DeepSeek-V4.1-Flash-1017 on Hugging Face
• deepseek-ai/deepseek-recipe on GitHub

Related stories

• China’s AI Stack Goes Vertical: How DeepSeek V4 Just Changed the Game
• DeepSeek Is Reportedly Close to Raising About 100 Billion Yuan…
• Alibaba’s MAI-UI Team Open-Sources OpenGCUA for Generalist Computer-Use Agents