DeepSeek V4.1 Flash 1017: What Changed, Benchmarks & Cost
DeepSeek-V4.1-Flash-1017 is DeepSeek's Oct 17, 2026 update to V4.1 Flash, and the model behind the deepseek-flash API name. DeepSeek reports higher scores than V4.1 Flash on all 11 benchmarks it lists, with shorter outputs. DeepSeek says API prices didn't change; we'll publish them here once we've checked them against its pricing page.
01 — What changed
What changed in DeepSeek V4.1 Flash 1017
A minor update, not a new model
DeepSeek calls 1017 a minor update to V4.1 Flash, with gains in reasoning, coding, agent and visual benchmarks.
Source: Hugging Face model card
Trained to finish whole projects
DeepSeek used reinforcement learning with self-feedback to improve end-to-end project work. Its examples include a web racing game and a 3D model of the Forbidden City.
Source: DeepSeek release note
Faster, with shorter outputs
DeepSeek says 1017 completes tasks faster and writes shorter outputs than V4.1 Flash, so the same work bills fewer output tokens.
Source: DeepSeek change log
Tuned for coding agents
Optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. DeepSeek recommends DeepSeek Harness 0.2.1-rc.1 or later.
Source: DeepSeek change log
Same API name, same price
Keep calling deepseek-flash. Requests to the retired names deepseek-v4-flash and deepseek-v4-flash-vision-exp are served by 1017 too, at the Flash price.
Source: DeepSeek pricing page
Open weights
The weights are on Hugging Face and ModelScope under the MIT license.
Source: DeepSeek release note
Not new in 1017 · from the V4.1 Flash release, Sep 10
- Native image understanding
- The 1M-token context window
- The architecture: a 552B-parameter mixture of experts with 8B active parameters for input and 16B for output
- The lower API prices
02 — Benchmarks
DeepSeek V4.1 Flash 1017 vs V4.1 Flash benchmarks
| Benchmark | V4.1 Flash | V4.1 Flash 1017 | Change |
|---|---|---|---|
| HLENo tools · [text-only subset] | 36.8 [39.1] | 42.6 [44.0] | +5.8 |
| Terminal-Bench 4.0Claude Code harness | 31.2 | 42.3 | +11.1 |
| DeepSWE v1.1Mini-SWE-V2 harness | 74.2 | 77.0 | +2.8 |
| ProgramBench | 20.3 | 29.1 | +8.8 |
| NL2Repo-Bench | 65.4 | 68.2 | +2.8 |
| Automation-Bench v1.0.6 | 54.8 | 60.5 | +5.7 |
| Agents' Last Exam | 31.8 | 35.4 | +3.6 |
| Apex Bench v1.1 | 64.6 | 72.1 | +7.5 |
| ChartographyWith tools | 78.9 | 83.0 | +4.1 |
| BabyVisionWith tools | 89.6 | 92.5 | +2.9 |
| ZeroBench-mainWith tools · Pass@5 | 49.0 | 51.0 | +2.0 |
- All scores are DeepSeek's own, from its model card. No independent results had been published at release (The Decoder, Oct 17).
- HLE ran without tools; the bracketed score is its text-only subset. Terminal-Bench 4.0 ran in the Claude Code harness and DeepSWE v1.1 in Mini-SWE-V2. ZeroBench-main is Pass@5.
- DeepSeek's table also lists Kimi K3, GLM 5.3, GPT-5.6 Sol and Claude Opus 5 — not GPT-6.1 Sol or Claude Opus 5.5, the current models we price. We link to it rather than copy those columns.
Cost per task · read from DeepSeek’s chart
DeepSeek's DeepSWE v1.1 chart (mini-swe-v2 harness) puts 1017 at 77.0% at max effort for an average of a little under $0.20 a task, against 74.2% for about $0.50 with V4.1 Flash. We read the costs off a log-scale chart, so treat them as approximate. Prices didn't change, so a lower cost per task means each task billed fewer tokens, or cheaper ones such as cache hits.
03 — API pricing
DeepSeek V4.1 Flash API pricing: peak and off-peak
Unverified · price pending review
We publish DeepSeek's prices after checking them against its pricing page, so there are no numbers here yet.
DeepSeek’s Models & Pricing page ↗04 — Cost calculator
Cost calculator
Unverified · price pending review
The calculator opens once DeepSeek's prices pass our review. We don't estimate from prices we haven't checked.
DeepSeek’s Models & Pricing page ↗05 — Use the API
How to use DeepSeek V4.1 Flash 1017 through the API
- OpenAI formatbase_url
- https://api.deepseek.com
- Anthropic formatbase_url
- https://api.deepseek.com/anthropic
- Model name
- deepseek-flash
- Also accepts
- Responses API
Quick start
Languagecurl https://api.deepseek.com/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $DEEPSEEK_API_KEY" \
-d '{
"model": "deepseek-flash",
"messages": [{"role": "user", "content": "Hello"}],
"thinking": {"type": "enabled"},
"reasoning_effort": "high"
}'Thinking mode switches
| Request format | On / off | Effort |
|---|---|---|
| OpenAI (Chat Completions) | "thinking": {"type": "enabled" | "disabled"} | "reasoning_effort": "low" | "high" | "max" |
| Anthropic | "thinking": {"type": "enabled" | "disabled"} | "output_config": {"effort": "low" | "high" | "max"} |
| Responses API | "reasoning": {"effort": "none"} turns it off | "reasoning": {"effort": "low" | "high" | "max"} |
Thinking is on by default at effort high. In thinking mode, temperature and the penalty parameters have no effect. The reasoning comes back as reasoning_content, and its tokens are counted in completion_tokens, so they bill as output. DeepSeek’s Thinking Mode guide.
Coding agents · DeepSeek’s setup guides
Rather chat than code? DeepSeek’s website and app already run 1017: chat.deepseek.com ↗
06 — Hugging Face
Run DeepSeek V4.1 Flash 1017 from Hugging Face
Download the weights
About 510 GB in 48 files. The same repository is mirrored on ModelScope.
Encode prompts with DeepSeek's code
There's no Jinja chat template. Use encoding/encoding.py from the repository, which ships with test cases, or deepseek-recipe, DeepSeek's Rust libraries with Python bindings, in production.
Run inference
The inference/ folder has the weight conversion script (convert.py), generate.py and run.sh. DeepSeek recommends temperature 1.0, top_p 1.0 and max_tokens of 256K or more.
Download with the Hugging Face CLI
hf download deepseek-ai/DeepSeek-V4.1-Flash-1017 --local-dir DeepSeek-V4.1-Flash-1017- Repository
- deepseek-ai/DeepSeek-V4.1-Flash-1017
- License
- MIT, code and weights
- Weights
- 48 safetensors files · about 510 GB
- Precision
- FP8, with FP4 expert weights
- Chat template
- No Jinja template
- Local inference
- inference/ folder
- Recommended sampling
- temperature 1.0 · top_p 1.0
- Recommended max_tokens
- 256K or more
- Tokenizer
- tokenizer.json · 6.4 MB
- Hosted on Hugging Face
- No inference provider (Oct 18)
07 — FAQ
DeepSeek V4.1 Flash 1017 FAQ
What is DeepSeek V4.1 Flash 1017?
DeepSeek-V4.1-Flash-1017 is the Oct 17, 2026 update to DeepSeek V4.1 Flash, which launched Sep 10, 2026. It answers the deepseek-flash model name on the DeepSeek API and runs on DeepSeek's website and app. Its weights are on Hugging Face and ModelScope under the MIT license.
What changed from DeepSeek V4.1 Flash?
DeepSeek reports higher scores on all 11 benchmarks it lists — Terminal-Bench 4.0 rose from 31.2 to 42.3 and ProgramBench rose from 20.3 to 29.1 — and says 1017 finishes tasks faster with shorter outputs. It was tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. Image understanding and the 1M-token context were already in V4.1 Flash.
How much does the DeepSeek V4.1 Flash 1017 API cost?
DeepSeek says the 1017 update didn't change API prices. We'll publish its rates here once we've checked them against DeepSeek's pricing page.
When are DeepSeek's off-peak hours?
We'll publish DeepSeek's peak and off-peak schedule with its prices once we've checked them against DeepSeek's pricing page.
Do I need to change my code to use 1017?
No. The deepseek-flash model name now serves 1017, and requests to the retired names deepseek-v4-flash and deepseek-v4-flash-vision-exp are served by 1017 at the Flash price. DeepSeek's API has no dated model name, so deepseek-flash will move again with the next update; to stay on 1017, use a provider that lists the snapshot, such as Vercel AI Gateway's deepseek/deepseek-v4.1-flash-1017, or run the weights yourself.
Is DeepSeek V4.1 Flash 1017 open source, and can I run it locally?
Its weights are open: deepseek-ai/DeepSeek-V4.1-Flash-1017 on Hugging Face and ModelScope, under the MIT license. They take about 510 GB across 48 files, and DeepSeek describes V4.1 Flash as a 552B-parameter mixture of experts, so plan for a multi-GPU server, not a laptop. There's no Jinja chat template; DeepSeek ships encoding/encoding.py and the deepseek-recipe libraries instead.
Can I use DeepSeek V4.1 Flash 1017 without the API?
Yes. DeepSeek switched its website and app to 1017 on Oct 17, 2026, so you can chat with it there without an API key. Use the API to build it into your own app or agent.
Sources
Sources & change log
Official pages · checked Oct 18, 2026
- DeepSeekChange Log
- DeepSeekDeepSeek-V4.1-Flash-1017 release, Oct 17, 2026
- DeepSeekModels & Pricing
- DeepSeekCreate Chat Completion
- DeepSeekThinking Mode
- DeepSeekIntroducing DeepSeek-V4.1-Flash, Sep 10, 2026
- Hugging Facedeepseek-ai/DeepSeek-V4.1-Flash-1017 model card
- ModelScopedeepseek-ai/DeepSeek-V4.1-Flash-1017 mirror
Elsewhere · snapshot names and release coverage
- VercelDeepSeek V4.1 Flash 1017 on AI Gateway
- DeepInfraDeepSeek V4.1 Flash 1017 API
- The DecoderRelease report, Oct 17, 2026
Change log
- Page published for the 1017 snapshot. Benchmarks from DeepSeek’s model card; prices pending review.
Pilot
Already paying an LLM API bill?
Join the pilot for project cost reviews and budgets. Two questions first — we only ask for your email if it fits.