DeepSeek V4.1 Flash is the open weights model that cut the gap between the best Chinese and US AI models on the LiveBench benchmark to about 3%, according to Bloomberg Intelligence. Released on September 10, 2026 under the MIT license, it scored 81.1 against 83.4 for Anthropic’s top entry in an October 4 leaderboard snapshot, while costing as little as $0.15 per million input tokens. A refreshed checkpoint, V4.1 Flash 1017, followed on October 17.
Key takeaways
- LiveBench overall: 81.1 for DeepSeek V4.1 Flash versus 83.4 for Anthropic’s leading model, a 2.3 point gap (about 3%)
- Trend: Bloomberg Intelligence put the US to China gap at about 15% earlier in 2026 and about 9% in May
- Agentic coding: DeepSeek 77.3 versus 66.1 for the Anthropic entry in the same snapshot
- Price: $0.30 input and $1.20 output per million tokens at peak, half that off peak
- License: MIT, with weights on Hugging Face; 1 million token context and native image input
What DeepSeek V4.1 Flash is
V4.1 Flash is the lighter, faster member of DeepSeek’s V4 family, which launched in April 2026. Despite the “Flash” label it is not small. According to The Next Web’s reading of the model card, it has 552 billion parameters in total, but only a fraction work on any single token: about 8 billion when reading input and 16 billion when writing output. That sparse design is how DeepSeek keeps serving costs low. DeepSeek calls the architecture a “causal encoder decoder” and says the model was pretrained on 45 trillion tokens.
Three practical features stand out. The context window reaches 1 million tokens, which is enough for large codebases or long document sets. The model understands images natively, though DeepSeek’s own technical report suggests it still lags the best closed systems on complicated images. And it needs far less memory per token for its conversation cache than V4 Flash did, which makes long sessions cheaper to run.
| Spec | DeepSeek V4.1 Flash |
|---|---|
| Release date | September 10, 2026 (checkpoint 1017 on October 17, 2026) |
| Parameters | 552 billion total; about 8 billion active for input, 16 billion for output (September model card) |
| Context window | Up to 1 million tokens, 384K max output on the API |
| Input types | Text and images |
| License | MIT, open weights on Hugging Face |
| API name | deepseek-flash |
The LiveBench numbers in context
LiveBench refreshes its questions regularly to reduce the chance that models have seen the answers during training, which is why analysts like it for comparisons across labs. Robert Lea, a senior analyst at Bloomberg Intelligence, used the October 4 snapshot to argue that the gap between the top US and Chinese models had hit a record low.
| Measure | DeepSeek V4.1 Flash | Top US entry (Anthropic) |
|---|---|---|
| LiveBench overall (Oct 4 snapshot) | 81.1 | 83.4 |
| Agentic coding sub score | 77.3 | 66.1 |
| Gap to the US leader | About 3% | n/a |
| Gap earlier in 2026 | About 15% | n/a |
| Gap in May 2026 | About 9% | n/a |
The agentic coding result is the eye catcher. That category measures whether a model can plan, write and fix code across several steps, which is close to what developers use AI agents for every day. It is one of the first times a Chinese open weights model has led a US flagship in a category that maps so directly to paid work.
Keep the limits in mind. A single leaderboard snapshot moves week to week. AI Weekly notes that only three of the top 15 LiveBench entries were Chinese, so the catch up is concentrated at the very top rather than across the field. And DeepSeek’s own launch figures were mixed: on its scorecard, V4.1 Flash edged Claude Opus 5 and GPT 5.6 Sol on DeepSWE (74.2 versus 74.0 and 73.0) but trailed Opus 5 badly on Humanity’s Last Exam (36.8 versus 56.3) and ProgramBench (20.3 versus 37.0). Those are company reported numbers, not independent tests.
Pricing: why builders are paying attention
Price is where V4.1 Flash puts real pressure on rivals. DeepSeek charges different rates at peak and off peak hours, with off peak rates exactly half.
| Per 1 million tokens | Off peak | Peak |
|---|---|---|
| Input, cache hit | $0.003 | $0.006 |
| Input, cache miss | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. For US users that peak window falls mostly overnight and in the early morning Eastern time, so a lot of normal US daytime traffic is billed at the off peak rate. Bloomberg Intelligence estimated the September repricing cut costs by up to 32%, after DeepSeek had raised prices in August. Lea described the move as “a more rational approach to pricing,” and he does not expect China’s AI sector to be profitable before 2030 because of a domestic price war.
For agent builders this matters more than headline benchmark points. Agents burn through tokens, so a model that is close to the frontier and costs a fraction per task is exactly what teams switch to. Implicator reported that in the week to September 7, Chinese models on OpenRouter processed more tokens than US models, even though US models still handled slightly more requests. The money question is now central to the US side too, as the debate over when Anthropic will go public shows.
Why the gap is closing
Lea credits two things: Chinese labs getting better technically, and tuning their models for domestic hardware. Huawei said in April that its full Ascend SuperNode lineup was adapted to DeepSeek V4 from day zero, which lets Chinese customers run frontier class inference without export controlled Nvidia chips. That is also why China is pouring money into projects like the Wuhu AI data center cluster and homegrown accelerators such as Lisuan’s 24GB workstation GPU.
US officials offer a competing explanation. On September 8, 2026, the NSA, FBI and CISA named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI as running “industrial scale” distillation campaigns against US frontier models such as Claude, GPT, Gemini and Grok since at least late 2024. Distillation means training a model on another model’s outputs. The LiveBench numbers neither prove nor disprove that claim; they only show the result.
What to keep in perspective
- One benchmark is not the whole story. LiveBench does not measure reliability, safety testing or enterprise support
- The lead is narrow. Only a few Chinese entries sit near the top of the leaderboard
- Supply chains still matter. Chip access, energy and geopolitics can shift the race quickly, as the Iran war’s risk to chip supplies showed
- Self reported scores need checking. Most launch benchmarks came from DeepSeek itself
Latest update, November 2026
Two things have happened since the Bloomberg Intelligence note. On October 17, DeepSeek released V4.1 Flash 1017, a checkpoint update under the same MIT license and the same deepseek-flash API name. DeepSeek reports 77.0 on DeepSWE for the new checkpoint, ahead of Claude Opus 5 and GPT 5.6 Sol on its own scorecard, and says it uses shorter answers that cut cost per task. The Hugging Face listing shows 763 billion parameters, a figure DeepSeek has not reconciled with the 552 billion in September. DeepSeek’s pricing page now serves the 1017 checkpoint under deepseek-flash at the rates above, and still lists V4 Pro as a separate, more expensive model.
Then on October 20, the Department of Homeland Security published an order requiring federal civilian agencies to stop buying DeepSeek products and remove them, including third party products built on DeepSeek technology. That does not affect private companies or individuals, but it does mean any vendor selling to civilian federal agencies cannot ship V4.1 Flash inside its product.
What it means for you
- If you build with AI, V4.1 Flash is worth testing on coding and agent workloads, especially if you can schedule heavy jobs in off peak hours.
- If data location worries you, remember the hosted API runs in China. The MIT weights let you self host instead, but you still need your own security review.
- If you sell to the US government, check your contracts: DeepSeek is now excluded for civilian agencies, and defense contractors faced restrictions earlier under the FY2026 defense bill.
- If you follow the industry, expect the shrinking gap to feature in every chip export debate and in pricing decisions at US labs.
A 3% gap does not mean the race is tied. It does mean the old assumption, that US labs hold a safe lead and can price accordingly, no longer holds without checking the leaderboard first.
FAQ
What is DeepSeek V4.1 Flash?
It is an open weights large language model from Chinese lab DeepSeek, released on September 10, 2026 under the MIT license. It supports text and image input, a context window of up to 1 million tokens and is served on DeepSeek’s API as deepseek-flash.
How close is DeepSeek V4.1 Flash to the best US model?
In the LiveBench snapshot of October 4, 2026, it scored 81.1 against 83.4 for Anthropic’s leading entry, a gap Bloomberg Intelligence put at about 3%. That gap was about 15% earlier in 2026 and about 9% in May.
How much does DeepSeek V4.1 Flash cost?
On DeepSeek’s API, input costs $0.30 per million tokens at peak and $0.15 off peak, and output costs $1.20 at peak and $0.60 off peak. Cached input is far cheaper, at $0.006 peak and $0.003 off peak.
Is DeepSeek V4.1 Flash free to download?
Yes. The weights are published on Hugging Face under the MIT license, so anyone can download, modify and run the model on their own hardware.
What is DeepSeek V4.1 Flash 1017?
It is a checkpoint update released on October 17, 2026. It keeps the MIT license and the deepseek-flash API name, and DeepSeek says it scores higher on coding tests while using shorter answers.
Can US government agencies use DeepSeek V4.1 Flash?
Federal civilian agencies cannot. A Department of Homeland Security order published on October 20, 2026 bars them from buying DeepSeek products and requires removal, including third party products built on DeepSeek technology.
Sources
- Implicator: DeepSeek cuts US AI benchmark lead to about 3%
- AI Weekly: DeepSeek V4.1 Flash narrows US China AI gap to 3% on LiveBench
- The Next Web: DeepSeek launches V4.1 Flash
- DeepSeek API Docs: Models and pricing
- Let’s Data Science: DeepSeek’s new model on Saturday, DHS ban on Monday
- TechRepublic: DeepSeek V4 AI model backed by Huawei Ascend chips
- ExecutiveGov: NSA, FBI, CISA warn of China AI distillation attacks
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.

