DeepSeek silently swapped its Flash API model; new checkpoint reports benchmark gains and cache savings

On October 17, DeepSeek replaced the model behind its deepseek-flash API endpoint without a formal changelog, so the same model name could route to a different checkpoint from one day to the next. The new checkpoint, DeepSeek-V4.1-Flash-1017, is published on Hugging Face under the MIT license. DeepSeek calls it a minor update, though its own benchmark table shows gains on all 11 reported tests. A commentator checked the public config file and computed KV cache memory at 1M tokens, finding that the widely circulated figure of 890 bytes per token depends heavily on context length and is therefore incomplete as a constant. The API pricing shows cached input reads priced 50 times below uncached input, which the author reads as consistent with the cache compression work. The vendor's own cost comparison claims roughly 10x to 50x savings per task, but these numbers rely on DeepSeek's own harness and remain unverified.