DeepSeek V4.1 Flash is DeepSeek's open-weights, MIT-licensed Flash model: a 552B-parameter mixture-of-experts model with a 1M-token context window, native image understanding, and API pricing of $0.30 per million input tokens and $1.20 per million output tokens at peak. The current build, DeepSeek-V4.1-Flash-1017, shipped on 17 October 2026 and now sits behind the deepseek-flash API name.
📺 Watch: Run DeepSeek V4.1 Flash Completely Free inside OpenCode!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
I covered it on 18 October in "DeepSeek V4.1 Flash 1017 Builds Apps in MINUTES!", then again on 23 October running it free inside OpenCode. This page is the model itself: what changed, the specs, the benchmarks, the price and how to get your hands on it.
What DeepSeek V4.1 Flash is in one minute
There are two releases to know about.
- DeepSeek-V4.1-Flash launched on 10 September 2026. DeepSeek called it "Smarter, Faster, More Efficient" and retired V4 Flash and V4 Flash Vision Exp the same day.
- DeepSeek-V4.1-Flash-1017 landed on 17 October 2026. The model card calls it a minor update with gains in reasoning, coding, agent tasks and visual benchmarks, plus shorter outputs and faster task completion.
DeepSeek describes the 1017 build as a Pareto improvement. Better scores and lower cost at the same time. It was trained with reinforcement learning using self-feedback, aimed at end-to-end project work, from games to 3D modeling. The launch post showed it building a 3D model of the Forbidden City, a browser racing game and a Portal 2 recreation in Godot.
One more thing most people miss. Since 14 September, every request sent to deepseek-v4-pro on DeepSeek's API is routed to V4.1 Flash and billed at the Flash price, until V4.1 Pro launches. So right now, Flash is effectively DeepSeek's top model on its own API.
If you want to put models like this to work instead of just reading about them, the AI Profit Boardroom ($69/mo) gives you my Agent OS, a 30-day AI agent roadmap, daily tutorials as new tools drop, four coaching calls a week and the prompt library. Over 3,400 members are in there.
What's new vs V4 Flash
This isn't a light refresh. V4.1 Flash is a new architecture. Per the Hugging Face model card, here's what changed from V4 Flash:
- Causal Encoder-Decoder design. 40 Transformer layers, split into a 20-layer encoder and a 20-layer decoder.
- Much smaller memory footprint. The persistent KV cache drops to about one eighth of V4 Flash, and FP4 caching brings the global KV cache to 890 bytes per token, about a quarter of V4 Flash.
- Adjustable reasoning effort. You can set effort on a scale from 1 to 100 instead of a few fixed levels.
- Native vision. A DeepSeek-ViT image encoder trained from scratch, so text and images go into the same model.
- Bigger pretraining run. 45T tokens, with context extended to 1M tokens.
Why should you care about KV cache numbers? Because long context is what agents burn. A smaller cache per token is a big part of how DeepSeek keeps a 1M-context model this cheap.
📺 Watch: Hermes + DeepSeek V4 Flash is WILD (FREE)
DeepSeek V4.1 Flash specs
The numbers that matter, from DeepSeek's launch post, pricing page and model cards:
- Size: 552B backbone parameters, with 8B active during prefill and 16B active during decode.
- Experts: 384 routed experts plus 1 shared expert per MoE layer, 6 routed experts active per token.
- Context: 1M tokens. Max output on the API is 384K tokens.
- Modes: thinking (the default) and non-thinking.
- API features: JSON output, tool calls, vision, the Responses API, the Anthropic API format, chat prefix completion (beta) and FIM completion (non-thinking only).
- License: MIT.
- Recommended settings: temperature 1.0, top_p 1.0, and a max tokens setting of 256K or more, per the 1017 card.
That Anthropic API support is the quiet win. Tools built for Claude-style endpoints can talk to it without a rewrite.
📺 Watch: OpenCode + DeepSeek V4 Flash Trick Changes Everything! (FREE!)
Benchmarks
Here's the jump across all three versions on the agent and coding tests. The V4 Flash and September V4.1 Flash numbers come from the V4.1 Flash model card. The 1017 numbers and the Opus 5 column come from the 1017 model card.
| Benchmark | V4 Flash | V4.1 Flash (Sept) | V4.1 Flash 1017 | Opus 5 |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 7.0 | 31.2 | 42.3 | 51.8 |
| DeepSWE v1.1 | 54.4 | 74.2 | 77.0 | 74.0 |
| AutomationBench | 37.7 | 54.8 | 60.5 | 50.3 |
| Agents' Last Exam | 25.2 | 31.8 | 35.4 | 28.6 |
| HLE (no tools) | 37.8 | 36.8 | 42.6 | 56.3 |
Three things jump out.
First, Terminal-Bench 4.0 went from 7.0 to 42.3 in two releases. That's the test closest to "can it actually drive a terminal like a developer", and V4 Flash basically couldn't.
Second, on DeepSWE v1.1, AutomationBench and Agents' Last Exam, DeepSeek's card puts the 1017 build ahead of Opus 5. DeepSeek ran Terminal-Bench 4.0 in the Claude Code harness and DeepSWE in Mini-SWE-V2.
Third, it's not a clean sweep. Opus 5 still leads on Terminal-Bench 4.0 and by a wide margin on HLE. So this is a cheap workhorse for agent and coding jobs, not a replacement for the frontier on hard reasoning.
Benchmarks are the vendor's numbers. The real test is your workflow, which is why I score models on my own tasks in Goldie Bench.
Not sure where a model like this fits in your business? Grab a free AI strategy session and we'll map it out.
Price
DeepSeek's pricing page lists deepseek-flash per million tokens:
- Input, cache miss: $0.30 peak, $0.15 off-peak.
- Input, cache hit: $0.006 peak, $0.003 off-peak.
- Output: $1.20 peak, $0.60 off-peak.
Off-peak is half price. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else, including whole weekends, is off-peak.
So if you run batch jobs, schedule them outside those windows and you cut the bill in half. Cache hits are so cheap that long agent sessions re-reading the same files cost almost nothing on the input side.
And there's a free route. OpenCode Zen lists a deepseek-v4.1-flash-free model, free for a limited time, with a zero-retention provider that doesn't train on your data. I walk through that setup in my DeepSeek V4.1 Flash OpenCode guide.
Open weights
The weights are public under MIT. You'll find them on Hugging Face at deepseek-ai/DeepSeek-V4.1-Flash-1017 and on ModelScope.
The model card lists these ways to run it yourself:
- vLLM with vllm serve.
- SGLang with sglang.launch_server, including a Docker image.
- Docker Model Runner: docker model run hf.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
- Quantized builds for llama.cpp, Ollama and LM Studio, linked from the card's quantizations section.
Heads up: the card ships no Jinja chat template. DeepSeek points you to a Python reference in the encoding folder and its deepseek-recipe toolkit instead. And at 552B parameters, the full model is a server job, not a laptop job. For most people, the API or OpenCode is the way in.
How to access DeepSeek V4.1 Flash
Four routes, from easiest to most hands-on.
- DeepSeek's website and app. The 1017 build is live there now.
- The API. Call deepseek-flash. The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work and are served by V4.1-Flash-1017 at the Flash price.
- A coding agent. DeepSeek says 1017 is optimized for DeepSeek Harness (v0.2.1-rc.1 or later), Tencent CodeBuddy and WorkBuddy. My DeepSeek V4 Flash harness walkthrough covers the harness side. OpenCode works too, through its DeepSeek provider or the free Zen model.
- Self-host the open weights with vLLM, SGLang or a quantized build.
Want a second cheap open model to compare it with? Look at my GLM 5.3 Flash OpenCode setup. Both sit in OpenCode's free list right now.
Here's how I use it. I run Hermes every day inside my Agent OS with a profile per model, so a cheap, fast model like this gets the bulk work and the frontier model gets the hard calls. My Agent OS guide shows how that's laid out.
If you'd like help building that kind of setup for your own business, book a free AI strategy session.
FAQ
Is DeepSeek V4.1 Flash open source?
The weights are open under the MIT license, on Hugging Face and ModelScope. You can download, run and fine-tune it.
What's the difference between V4.1 Flash and V4.1 Flash 1017?
1017 is the 17 October 2026 update to the September release. Same family, better scores on every benchmark DeepSeek listed for both, shorter outputs and faster task completion. The API name deepseek-flash now serves 1017.
How much context does DeepSeek V4.1 Flash have?
1M tokens, with up to 384K tokens of output on DeepSeek's API.
Does DeepSeek V4.1 Flash support images?
Yes. It understands images natively through its DeepSeek-ViT encoder, and vision is listed as supported for deepseek-flash on the API.
Ready to turn cheap models into real agent workflows? Join the AI Profit Boardroom, or start with a free AI strategy session. That's DeepSeek V4.1 Flash.
