DeepSeek V4.1 Flash: Specs, Pricing, Benchmarks and the 1017 Update
DeepSeek V4.1 Flash is DeepSeek’s fast, low-cost model, and the October update made it even better: higher scores, faster jobs and shorter answers. It has a 1M-token context, open MIT weights and output pricing as low as $0.60 per million tokens off-peak. In the video below I run it free inside OpenCode. Here’s how it’s built, what the 1017 update changed, what it costs and how to use it.
Key takeaways
552B mixture-of-experts model with only 8B active for input and 16B for output.
The 1017 update lifted Terminal-Bench 4.0 from 31.2 to 42.3.
From $0.15 in / $0.60 out per million tokens off-peak.
DeepSeek launched V4.1 Flash on 10 September 2026. It’s a 552-billion-parameter mixture-of-experts model, but only a small part runs for each token: 8B active when reading input and 16B when writing output. That’s why it’s fast and cheap.
It also uses a new design DeepSeek calls a Causal Encoder-Decoder. The model card lists 40 layers (20 encoder and 20 decoder) and 384 routed experts plus one shared expert, with six active at a time. The practical win is memory: its KV cache needs about a quarter of the GPU memory of the previous generation.
DeepSeek V4.1 Flash at a glance
Total size
552B (MoE)
Active
8B input / 16B output
Context
1M tokens
Max output
384K tokens
Training
45T tokens
Licence
MIT
API name
deepseek-flash
It also understands images natively. I covered that in a separate post on DeepSeek V4.1 Flash vision, so I’ll keep this page on the model itself.
What the 1017 update changed
On 17 October 2026 DeepSeek shipped V4.1-Flash-1017. DeepSeek calls it a minor update, but the numbers moved a lot. It also finishes tasks faster and writes shorter answers, which saves you tokens.
Benchmark
V4.1 Flash
V4.1 Flash 1017
Terminal-Bench 4.0
31.2
42.3
DeepSWE
74.2
77.0
ProgramBench
20.3
29.1
AutomationBench
54.8
60.5
Agents’ Last Exam
31.8
35.4
HLE
36.8
42.6
Apex Bench
64.6
72.1
DeepSeek trained the update with reinforcement learning on full project builds, with the model checking its own work. It’s tuned for DeepSeek Harness (v0.2.1-rc.1 and later), Tencent CodeBuddy and WorkBuddy, but it works anywhere that speaks the API.
Want DeepSeek running cheap agent jobs for your business? Inside the AI Profit Boardroom you get the Agent OS — Hermes, OpenClaw, OpenCode, Claude and Codex in one place — plus step-by-step Hermes training and four live coaching calls a week.
Compared with the old V4 Flash, the September release was already a big step. Terminal-Bench 4.0 went from 7.0 to 31.2, DeepSWE from 54.4 to 74.2 and AutomationBench from 37.7 to 54.8. Then 1017 pushed it further.
For context, the 1017 model card compares it with Claude Opus 5. Opus 5 still leads on Terminal-Bench 4.0 (51.8 vs 42.3), but V4.1 Flash 1017 edges it on DeepSWE (77.0 vs 74.0) and AutomationBench (60.5 vs 50.3). That’s very strong for a model this cheap.
DeepSeek charges less outside its busy hours. From the pricing page, per million tokens:
Price
Peak
Off-peak
Input (cache hit)
$0.006
$0.003
Input (cache miss)
$0.30
$0.15
Output
$1.20
$0.60
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else is off-peak at half price. If you run big batch jobs, schedule them outside those windows.
DeepSeek also has a Search and Fetch API at $0.75 per 1,000 calls. Until 31 October 2026 at 16:00 UTC, it’s free when used through the server-side tools in DeepSeek Harness, Claude Code or Codex.
Model names and routing
The API name is deepseek-flash, which now points at V4.1-Flash-1017.
Old names deepseek-v4-flash and deepseek-v4-flash-vision-exp route to it. V4 Flash is retired.
Since 14 September, deepseek-v4-pro requests also go to V4.1 Flash at Flash prices until V4.1 Pro launches.
It supports thinking (the default) and non-thinking modes, JSON output, tool calls, the Responses API and the Anthropic API format.
The Anthropic-format support is handy. Tools built for Claude can often point at DeepSeek with only a base URL and key change.
How to run DeepSeek V4.1 Flash
You have three main options:
The DeepSeek API: easiest, and cheap off-peak.
Your own hardware: the 1017 weights are on Hugging Face and ModelScope. Serve them with vLLM, SGLang or Docker Model Runner. Community quantisations exist for llama.cpp, Ollama and LM Studio.
OpenCode: a free version is available there for a limited time. I wrote a full walkthrough: DeepSeek V4.1 Flash in OpenCode.
docker model run hf.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
DeepSeek recommends temperature 1.0, top_p 1.0 and a max token budget of at least 256K for hard tasks.
If you run lots of agent jobs and want them cheap, DeepSeek V4.1 Flash is hard to beat. It’s great for coding agents, automation and long documents. If you need the very best terminal performance, a top closed model still leads. And as with any model, check its work on anything that matters.
Tips to keep DeepSeek costs down
Run batch jobs off-peak. Everything outside the peak windows is half price.
Reuse the same prompt start. Cache hits cost $0.006 per million tokens at peak, against $0.30 for a miss.
Use non-thinking mode for simple jobs like formatting or short replies.
Cap output length. Max output is 384K tokens, so set a sensible limit for everyday jobs.
Done right, a whole day of agent work can cost less than a coffee. Just check the bill after your first week so there are no surprises.
Self-hosting is another option once your volume is high. The weights are free under MIT, and with only 16B active parameters per output token it’s far lighter to serve than its 552B size suggests. You still need serious GPU memory for the full model, so most people should start on the API.
My verdict on DeepSeek V4.1 Flash
DeepSeek V4.1 Flash gives you strong agent scores, a 1M-token context and open MIT weights for pennies. The 1017 update made it faster and better at the same price. If you want a cheap workhorse model for your agents, DeepSeek V4.1 Flash should be on your shortlist.
About Julian Goldie
I’m Julian Goldie — founder of Goldie Agency and the AI Profit Boardroom. I test new AI agents almost every day for 432K+ YouTube subscribers and run Hermes, OpenClaw, OpenCode, Claude and Codex side by side in my own Agent OS. Claim your free strategy session or join the AI Profit Boardroom.
FAQ: deepseek v4.1 flash
What is DeepSeek V4.1 Flash?
DeepSeek’s fast, low-cost model: a 552B mixture-of-experts with 8B active for input and 16B for output, a 1M-token context and open MIT weights.
What changed in DeepSeek V4.1 Flash 1017?
The 17 October update raised benchmark scores, finished tasks faster and wrote shorter answers. Terminal-Bench 4.0 went from 31.2 to 42.3.
How much does DeepSeek V4.1 Flash cost?
$0.30 input and $1.20 output per million tokens at peak, and half that off-peak. Cached input is $0.006 at peak.
Is DeepSeek V4.1 Flash open source?
Yes. The weights are on Hugging Face and ModelScope under the MIT licence.
What is the DeepSeek V4.1 Flash API model name?
deepseek-flash. The old deepseek-v4-flash names route to it.
Ready to put Hermes to work in your business? Join the AI Profit Boardroom for the Agent OS, setup training and daily updates.
