Credit: Matheus Bertelli from Pexels / Canva DeepSeek released V4.1-Flash on Thursday. It is a smaller model that the company says beats its own flagship on coding and agent tasks, at a lower price. The Hangzhou lab announced it in a thread on X and published the weights on Hugging Face under the MIT licence. That lets anyone download, modify and run it.
The launch comes with a notable retirement. From 04:00 UTC on 14 September, DeepSeek will send every request made to V4-Pro, its top model, to V4.1-Flash instead. Those requests will be billed at the cheaper Flash rates. That arrangement runs until a V4.1-Pro arrives.
DeepSeek gave no date for it. A big model that works like a small one V4.1-Flash has 552 billion parameters in total. It uses only a small slice of them for each piece of text. DeepSeek calls the design a “causal encoder-decoder” and describes it as the smallest model in a new architecture family.
It activates 8 billion parameters per token when reading input and 16 billion when writing output. That split is aimed at AI agents. An agent that calls tools repeatedly spends much of its time reading fresh input, so cheaper reading means cheaper agents. The model handles up to 1 million tokens of context and understands images natively, according to its model card.
DeepSeek says it pre-trained the model on 45 trillion tokens. The bigger engineering claim concerns memory. Models keep a cache of what they have already processed, known as the KV cache. It saves them recomputing everything at every step.
For long agent sessions, that cache eats expensive high-bandwidth memory. DeepSeek says V4.1-Flash needs 890 bytes per token for it. That is about a quarter of what V4-Flash needed, and roughly 437 times less than its first model in 2023. Where it wins, and where it does not DeepSeek’s own benchmark table puts V4.1-Flash level with the best closed models on some coding tests.
On DeepSWE v1.1, a software engineering benchmark, it scores 74.2. Anthropic’s Claude Opus 5 scores 74.0 and OpenAI’s GPT-5.6 Sol 73.0. On CyberGym, a cybersecurity test, it scores 88.1, ahead of every rival with a listed score. The same table shows clear gaps.
On Humanity’s Last Exam, a hard academic test, it scores 36.8 to Opus 5’s 56.3. On ProgramBench it reaches 20.3 against 37.0 for Opus 5. On Terminal-Bench 3.0 it trails both US models. These are DeepSeek’s figures.
TNW has not independently verified them. Its closest Chinese rival looks weaker by DeepSeek’s count. The table puts V4.1-Flash ahead of Moonshot’s Kimi K3 on every agent and coding test where both have scores. Kimi K3 still leads on GPQA Diamond and Humanity’s Last Exam.
In August, TNW covered an experimental multimodal model that DeepSeek pitched against Anthropic. V4.1-Flash replaces it. The Decoder’s Jonathan Kemper, who read the technical report, flagged a less flattering part of it. DeepSeek describes reward hacking during training, with agents occasionally using newly published vulnerabilities or wiping important system files in their test environments.
It also admits the model lags the best closed systems at reading complicated images. A price cut, a month after a price rise DeepSeek says the cheaper architecture lets it serve more users at lower cost. It says it is “passing the savings on” to customers. The new rates took effect at 04:00 UTC on Thursday.
Off-peak, one million tokens of output cost $0.60, and cached input costs $0.003. Peak rates, which apply on weekday mornings UTC, are double. Bloomberg Intelligence analysts put the cut at as much as 32%. The move reverses the direction of travel from August, when DeepSeek launched a Claude Code rival and quadrupled its prices.
Customers still on V4-Pro will now pay Flash rates for what DeepSeek says is a better model. Cheap Chinese models have squeezed rivals for months, in what analysts have called a ‘death zone’ for US model makers. In early August, TNW reported that V4-Flash was already the cheapest well-known model to run. Rivals and investors react The pressure landed on Chinese competitors first.
Shares in MiniMax and Z.ai fell more than 8% in Hong Kong on Thursday. Alibaba slid more than 2%, Bloomberg’s Saritha Rai reported. The timing matters for DeepSeek’s own plans. Reuters reported that the company has begun preparing for a listing on Shanghai’s STAR Market.
That follows the IPO preparations TNW covered in July. DeepSeek is also courting operators directly. It says it will work with the open-source community on inference support, and invited anyone planning a deployment of 2,000 GPUs with a storage cluster to get in touch. Coding tools WorkBuddy and OpenCode already support the model.
The open question is whether DeepSeek’s benchmarks hold up in independent testing, and what V4.1-Pro costs when it arrives. Until then, every V4-Pro customer becomes a V4.1-Flash customer on Monday.
