DeepSeek ships 552B V4.1-Flash, replaces V4-Pro with lower-cost multimodal model

I see the weights are already on Hugging Face, but 552B is nearly double the original 284B Flash, so calling it local-friendly feels strained. The 890-byte-per-token KV cache and vision support are interesting, though.

The extra 196B Engram memory explains much of the size increase and score gain. I am waiting for V4.1 Pro.

V4 Flash was mostly FP4 at about 160GB, while this release appears to be FP8 and about 510GB. That moves it from dual Spark or Strix Halo territory to third-party quantization and perhaps four machines, unless the Engram weights can stay on SSD.

Total parameter count is not the speed metric here. It activates 8B during prefill and 14B or 16B during decode rather than always 13B, so "Flash" still fits.

I have seen SSD offload hurt prompt processing on my Mac Studio, from 550 to 300 pp, and generation from 40 to 33 tg. I would be happy if this model can still hit 300 pp.

I expect SSD streaming to work reasonably well because its MoE is very sparse and the Engram weights can sit on disk. The smaller per-session KV cache may make slow, unattended near-frontier inference workable on consumer machines with fast storage.

Flash is about inference speed, not total weights. If DeepSeek gets 2 to 2.5 times the older Flash speed while improving capability, with roughly 250 to 400 tokens per second, that is a Flash model even if it no longer fits in 256GB RAM.

My early use says it is strong, cheaper, and a good backup when a primary subscription hits its limit. It also follows requests that US models often refuse, such as downloading a ROM I legally own.

I use models for automated nightly bug triage that groups unique error messages and searches for the cause. This model is good for that job.

I test models against codebases with out-of-bounds access bugs. US models often block even a real codebase with no harmful intent, while Chinese models tend to comply with vulnerability-analysis prompts.

The Causal Encoder-Decoder design is the interesting part: 20 causal encoder layers feed 20 decoder layers, and the decoder projects global KV state from the final encoder states. A 1 to 100 reasoning-effort setting also makes cost versus accuracy adjustable, though benchmarks still need real-world validation.

I think this is the first Flash release with multimodal support and it already feels better than V4 Pro. DeepSeek is discontinuing V4 Pro, so Flash may become the main series.