Officials called it a “minor update.” But after the community took it apart, they found it was clearly a major overhaul.
On October 17, DeepSeek released V4.1-Flash-1017. The model card said it would “shorten task time and reduce output length.” The API name and pricing stayed the same, and the weights were open-sourced under MIT.
One developer compared the new weights with the September version layer by layer and reportedly found that about 40% of routed experts, 30% of memory, and 50% of the vision-tower expert tensors had been rewritten. Another developer ran quantization on two DGX Sparks and reportedly got 56.8 tokens/second for text and 92.2 tokens/second for code with the TensorFold engine—nearly twice vLLM’s speed. The architecture hadn’t changed, but the scale of the changes was on the level of a “major version.”
The price war has entered a new dimension, too: according to an engineer’s tests, DeepSeek completed a real coding task for only about 4.5 cents, but burned through roughly 550,000 tokens (Opus used only 30,000). Tokens are so cheap you can “waste” them and still pay less. OpenRouter data shows it processed about 38.7T tokens over the past 7 days—three times as many as the runner-up.
The timing is intriguing: the day before the MIT-licensed weights were released (October 16), the Ministry of Commerce and the Ministry of Science and Technology published a draft export-control proposal for public comment. It proposed adding 40 technologies to the control list, 23 of them AI-related. Models with compute of 10²⁶ FLOPs, more than 2 trillion parameters, or a global top-ten ranking would have their weights, source code, and architecture brought under controls. APIs and SaaS would be unaffected, and already open-sourced models would not be covered retroactively. The public comment period runs until November 15, as confirmed by multiple media outlets.
I think this was the week the AI narrative of “openness vs. control” blew its opening whistle: for the first time, model weights were written into an export-control list. In the short term, the cheaper models get, the less it costs to run AI agents at scale—and the more fuel there’ll be for on-chain AI narratives (such as $TAO $RENDER ).
There’s still the mystery of V4.1-Pro: reportedly, some accounts spotted its hidden model name in early access as far back as late September. Prediction markets put the odds of a release before the end of October at around 16%. One thing to watch: how the draft is finalized on November 15.
This is not investment advice. DYOR; trading involves risk.
#DeepSeekReleasesUpdatedAIModel
#DeepSeek #AI
On October 17, DeepSeek released V4.1-Flash-1017. The model card said it would “shorten task time and reduce output length.” The API name and pricing stayed the same, and the weights were open-sourced under MIT.
One developer compared the new weights with the September version layer by layer and reportedly found that about 40% of routed experts, 30% of memory, and 50% of the vision-tower expert tensors had been rewritten. Another developer ran quantization on two DGX Sparks and reportedly got 56.8 tokens/second for text and 92.2 tokens/second for code with the TensorFold engine—nearly twice vLLM’s speed. The architecture hadn’t changed, but the scale of the changes was on the level of a “major version.”
The price war has entered a new dimension, too: according to an engineer’s tests, DeepSeek completed a real coding task for only about 4.5 cents, but burned through roughly 550,000 tokens (Opus used only 30,000). Tokens are so cheap you can “waste” them and still pay less. OpenRouter data shows it processed about 38.7T tokens over the past 7 days—three times as many as the runner-up.
The timing is intriguing: the day before the MIT-licensed weights were released (October 16), the Ministry of Commerce and the Ministry of Science and Technology published a draft export-control proposal for public comment. It proposed adding 40 technologies to the control list, 23 of them AI-related. Models with compute of 10²⁶ FLOPs, more than 2 trillion parameters, or a global top-ten ranking would have their weights, source code, and architecture brought under controls. APIs and SaaS would be unaffected, and already open-sourced models would not be covered retroactively. The public comment period runs until November 15, as confirmed by multiple media outlets.
I think this was the week the AI narrative of “openness vs. control” blew its opening whistle: for the first time, model weights were written into an export-control list. In the short term, the cheaper models get, the less it costs to run AI agents at scale—and the more fuel there’ll be for on-chain AI narratives (such as $TAO $RENDER ).
There’s still the mystery of V4.1-Pro: reportedly, some accounts spotted its hidden model name in early access as far back as late September. Prediction markets put the odds of a release before the end of October at around 16%. One thing to watch: how the draft is finalized on November 15.
This is not investment advice. DYOR; trading involves risk.
#DeepSeekReleasesUpdatedAIModel
#DeepSeek #AI