@Dinosaur_liu: A horror story for all Chinese and American model makers: DeepSeek V4 Flash has only 284b parameters, with a mere 13b active parameters
Summary
It is claimed that DeepSeek V4 Flash has only 284 billion parameters and only 13 billion active parameters, posing an efficiency shock to model manufacturers in China and the US.
View Cached Full Text
Cached at: 08/03/26, 09:38 AM
Here’s a horror story for all AI model vendors in China and the US:
DeepSeek V4 Flash has only 284b parameters.
Its activated parameters are a mere 13b. https://t.co/xGWUEQH1Qc
Similar Articles
@Fenng: Saw this piece written by a self-media account — 'The latest fourth-generation WeLM-80B now has only 80 billion total parameters, with 3 billion activated, an activation rate of just 3.75%. For comparison — DeepSeek-V4-Flash, the domestic representative of extreme cost-performance, has 284 billion total parameters, 13 billion activated, activation rate of 4.6%...'
Fenng shares a self-media comparison between the fourth-generation WeLM-80B (80B total params, 3B activated, 3.75% activation rate) and DeepSeek-V4-Flash (284B total, 13B activated, 4.6% activation rate), with a humorous comment.
@geekbb: Took a look — this image shows DeepSeek V4 Flash's price moving forward, which can keep a whole bunch alive again. Fortunes turn; today it's the LLM vendors' turn to hail Liang as 'Saint Liang'
The author comments on DeepSeek V4 Flash's price cut, believing it can help more vendors survive, and jokingly says that LLM vendors should call Liang Wenfeng 'Saint Liang'.
@FuckAnthropic: Conducted a comparative analysis. Overall, DeepSeek V4 Flash-0731 is roughly a model at the level between Opus 4.7 and 4.8, entering the frontier Agent model competition with a minimal activation scale, and at about 1/12 to 1/60 of the token cost to enter the frontier Ag…
The author's comparative analysis concludes that DeepSeek V4 Flash-0731 achieves Opus 4.7–4.8 level performance with an extremely small activation scale, entering the frontier agent model tier at a very low token cost. It surpasses GLM-5.2 overall, but its shortfalls remain difficult repository-level coding and long-horizon engineering.
@Saccc_c: Want to add multimodal capabilities to DeepSeek v4 flash? I strongly recommend using it with qwen3.7-flash — currently the best value model combination. qwen3.7-flash is a lightweight multimodal model that is fast and well-suited to most image understanding tasks. Key point: the price is low enough, and new registrations get 100…
Recommends pairing DeepSeek v4 flash with qwen3.7-flash to add multimodal image understanding capabilities to DeepSeek at low cost, and provides simple configuration steps using Alibaba Cloud Bailian and Codex.
DeepSeek-V4-Flash-0731
DeepSeek announces DeepSeek-V4-Flash-0731, a frontier agent intelligence model positioned as offering advanced capabilities at Flash-level pricing.