DeepSeek-V4.1-Flash-1017 update: big agentic-coding gains, same price
Confirmed
The takeaway
On Oct 17, 2026 DeepSeek replaced V4.1-Flash on its web app, mobile app and API (still called deepseek-flash, prices unchanged) with DeepSeek-V4.1-Flash-1017, a “minor update” with open MIT-licensed weights on Hugging Face.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 17 October 2026
Your AI and this story
- GPT-6 Astra170 days after its cutoff
- Claude Opus 5.5109 days after its cutoff
- Gemini 4 Argon200 days after its cutoff
- Grok 4.7139 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 109 days before it.
Key facts
- Announced Oct 17, 2026 on DeepSeek’s API news page (‘DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real’) and on X (@deepseek_ai, ~872k views and ~12.9k likes by 13:00 CEST)
- Availability: web, app and API; ‘The API calling method and pricing remain unchanged — simply use deepseek-flash’ (API changelog). OpenRouter lists deepseek/deepseek-v4.1-flash-1017 at $0.15 / $0.60 per 1M input/output tokens with a 1,048,576-token context
- Open weights: huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017, MIT license, created Oct 17 02:40 UTC; Hugging Face’s safetensors metadata counts about 763B stored parameters (mostly FP8/INT8 tensors); recommended temperature 1.0, top_p 1.0, max_tokens ≥ 256K
- Official scores, 1017 vs the Sept 10 V4.1-Flash: HLE w/o tools 42.6 (text subset 44.0) vs 36.8 (39.1); Terminal-Bench 4.0 42.3 vs 31.2 (Claude Code harness); DeepSWE v1.1 77.0 vs 74.2 (mini-SWE-v2 harness); ProgramBench 29.1 vs 20.3; NL2Repo-Bench 68.2 vs 65.4; Automation-Bench 60.5 vs 54.8; Agents’ Last Exam 35.4 vs 31.8; Apex Bench v1.1 72.1 vs 64.6; Chartography 83.0, BabyVision 92.5, ZeroBench-main 51.0 (pass@5)
- Same table’s comparison columns (DeepSeek’s numbers): GPT-5.6 Sol HLE 44.5, Terminal-Bench 4.0 39.9, DeepSWE 73.0; Claude Opus 5 HLE 56.3, Terminal-Bench 4.0 51.8, ProgramBench 37.0; Kimi K3 and GLM-5.3 trail on most rows
- Training note (news page): ‘uses reinforcement learning with self-feedback to improve end-to-end project development’, with demos of a 3D Forbidden City model, a web racing game, a Portal 2 recreation in Godot and a 3D ‘Valley of Ages’ scene
- Tuned for DeepSeek Harness (v0.2.1-rc.1 or later), Tencent CodeBuddy and WorkBuddy; no Jinja chat template, instead an encoding reference implementation and the Rust ‘deepseek-recipe’ library that converts Messages, Chat Completions and Responses API requests into DeepSeek prompts
- Pandaily (Oct 17) reported that the ‘1017’ snapshot surfaced briefly, unannounced, alongside the DeepSeek Harness v0.2.1-rc1 release before the launch post
Show 1 more
- Hacker News: two front-page threads on Oct 17 (Hugging Face page 180 points; OpenRouter page 168 points)
What happened
Five weeks after V4.1-Flash (Sept 10), DeepSeek shipped a dated snapshot, DeepSeek-V4.1-Flash-1017, on Saturday Oct 17, 2026. It
became the default model on DeepSeek’s website, app and API under the same deepseek-flash name and price. The weights are on Hugging
Face under the MIT license. DeepSeek calls it a “minor update”, but the gains on its agentic benchmarks are large: Terminal-Bench 4.0 rose
by 11 points, ProgramBench by 9 and Humanity’s Last Exam (no tools) by about 6. DeepSeek also says the model finishes tasks faster and with
shorter outputs, which it calls a Pareto improvement in performance and cost. The same post opened the Search and Fetch APIs to everyone
(covered in DeepSeek Search and Fetch APIs for agents).
All benchmark numbers come from DeepSeek. Several of them (Terminal-Bench 4.0, ProgramBench, Agents’ Last Exam) are newer benchmarks with few independent results yet. The comparison columns use GPT-5.6 Sol and Claude Opus 5, not OpenAI’s and Anthropic’s current flagships (GPT-6 Astra, Claude Opus 5.5).
Why it matters
V4.1-Flash was already the open-weight model many developers used as a cheap coding agent. This update keeps the price at $0.15/$0.60 per million tokens and, by DeepSeek’s numbers, beats GPT-5.6 Sol on most agent benchmarks it lists. That widens the price gap between open Chinese models and the US frontier for agentic coding.
Sources
8 sources from 6 sites. Numbers match the chips in the text.
8 sources: 5 primary, 1 press, 2 reactions
Primary
- DeepSeek API news: DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real (2026-10-17)api-docs.deepseek.com, official
- DeepSeek API change log (V4.1-Flash-1017 update, pricing unchanged)api-docs.deepseek.com, official
- DeepSeek on X: DeepSeek-V4.1-Flash-1017 is here!x.com, official
- Hugging Face: deepseek-ai/DeepSeek-V4.1-Flash-1017 (MIT)huggingface.co, code
- OpenRouter: deepseek/deepseek-v4.1-flash-1017openrouter.ai, docs
Press
- Pandaily: DeepSeek Harness v0.2.1-rc1 arrives as an unannounced V4.1 Flash snapshot briefly surfacespandaily.com, press
Reactions
- Hacker News: DeepSeek-v4.1-Flash-1017 (180 points)news.ycombinator.com, discussion
- Hacker News: DeepSeek v4.1 Flash 1017 on OpenRouter (168 points)news.ycombinator.com, discussion
Changes
- Filed (read DeepSeek’s news page, change log and Hugging Face card, OpenRouter model list and the X post via fxtwitter)