DeepSeek's V4.1 Flash tops Kimi K3 on Terminal-Bench
DeepSeek says the 552-billion-parameter V4.1 Flash beats its own V4 Pro flagship while running cheaper, with only one benchmark made public.
In short
DeepSeek's V4.1 Flash is a cheaper, faster model that the company scores at 90.6 on Terminal-Bench 2.1, ahead of GPT-5.6 Sol (88.8), Kimi K3 (88.3) and DeepSeek's own V4 Pro (87.9).
At a glance
- Released September 10, 2026; DeepSeek describes a Causal-Encoder-Decoder architecture paired with Mixture-of-Experts (MoE).
- 552 billion total parameters: 8 billion active for input processing, 16 billion for response generation.
- Terminal-Bench 2.1 scores: V4.1 Flash 90.6, GPT-5.6 Sol 88.8, Kimi K3 88.3, DeepSeek V4 Pro 87.9.
- Native multimodal visual understanding is listed among the model's capabilities.
- No token pricing, no quantified speed gain and no weight-licence details appear in the report.
DeepSeek released V4.1 Flash on Thursday and says it edges past Moonshot AI's Kimi K3. On Terminal-Bench 2.1 — the only test named in the report — the company puts the model at 90.6. That compares with 88.8 for OpenAI's GPT-5.6 Sol, 88.3 for Kimi K3 and 87.9 for DeepSeek's own V4 Pro.
How the model is built
DeepSeek pairs a Causal-Encoder-Decoder layout with a Mixture-of-Experts design. Of 552 billion parameters in total, 8 billion are active while the model reads an input and 16 billion while it writes a response. Splitting those two loads is the mechanism behind the cheaper-and-faster claim. Native multimodal visual understanding is listed among the capabilities as well.
One benchmark, four models
Every figure here comes from DeepSeek, and no independent replication has been published. Beyond Terminal-Bench 2.1 the report names only categories — coding, cybersecurity and autonomous agent work — with no further test names and no per-test scores. The four models sit within a few points of one another on the same scale. A single published number is a thin basis for a leadership claim.
What DeepSeek did not say
The report gives no price per million tokens and no figure for how much faster V4.1 Flash runs than V4 Pro. It also does not state whether the weights are downloadable, or under what licence. The cost argument therefore arrives without a number attached to it.
The setting
Chinese developers are building under export limits on foreign chips and rising hardware bills. Efficiency there is a constraint, not a slogan. V4.1 Flash lands squarely inside that price-and-performance contest.
FAQ
What is DeepSeek V4.1 Flash?
A 552-billion-parameter model DeepSeek released on September 10, 2026, which the company says runs cheaper and faster than its V4 Pro flagship.
Does V4.1 Flash beat Kimi K3?
On Terminal-Bench 2.1 DeepSeek reports 90.6 for V4.1 Flash and 88.3 for Moonshot AI's Kimi K3. Those scores are self-reported and have not been independently verified.
Are the V4.1 Flash weights open?
The report does not say whether the weights can be downloaded, or under what licence, so any open-weights assumption is unconfirmed.
