DeepSeek V4.1-Flash: builders argue over the 890-byte KV cache

Builders spent Thursday arguing about one number: 890 bytes. That is DeepSeek's claimed KV-cache footprint per token in DeepSeek-V4.1-Flash, the open-weight model it dropped on Hugging Face under MIT — and the main announcement thread pulled roughly 1.37 million views in five hours.

Key facts

• The DeepSeek-V4.1-Flash repo is public, MIT-licensed, multimodal (image-text-to-text), and was last modified September 10, 2026 at 08:18 UTC — it passed 1,400 likes within a day.
• Architecture per the model card: a 552B-parameter MoE with a causal encoder-decoder design, 8B active parameters on prefill, 16B on decode, 1M-token context.
• API pricing per DeepSeek's pricing page: $0.003 per million cached input tokens off-peak, $0.15 cache-miss input off-peak, $0.60 output off-peak — peak rates double.
• The @deepseek_ai announcement thread drew ~14,374 likes, ~1,711 reposts, and ~1.37M views within five hours (point-in-time Grok x_search pull, September 10).

What's driving the conversation

The discourse has two camps. The amplifiers — including @jun_song and @UnslothAI — frame V4.1-Flash as the moment an open-weights model matched closed flagships while gutting serving costs. The claim doing the work: benchmarks ahead of DeepSeek-V4-Pro and competitive with Opus 5 and GPT-5.6-Sol on evals including DeepSWE, at a quarter of the prior generation's KV-cache footprint. Those are DeepSeek's numbers, repeated by fans.

The skeptics are running their own tests and posting early results suggesting the real-world gap is wider than the eval table. That is a normal release-week pattern — but worth naming, because the "open mogged closed" framing traveled far faster than any replicated result. A third thread, from @tianyi, matters more to builders than the benchmark fight: DeepSeek staff posted new GitHub deployment repos (deepseek-recipe, DeepSelect, DeepJIT), the practical plumbing for actually serving this thing.

The substance

Verified, because we checked the primary sources: the weights exist, they are MIT-licensed, the repo is public and multimodal. The model card's architecture claims are specific enough to be falsifiable — a 40-layer transformer split into a 20-layer causal encoder and 20-layer decoder, FP4 KV caching, sliding-window bounded replay cutting persistent KV cache to roughly an eighth of DeepSeek-V4-Flash's. Verified also: the pricing page, which confirms deepseek-flash now serves V4.1-Flash, and that from 12:00 Beijing time on September 14, deepseek-v4-pro requests route to V4.1-Flash at Flash prices until V4.1-Pro ships.

Claims, not verified: every benchmark number, the "comprehensively surpassed V4 Pro" line on the pricing page, and the efficiency figures. Third-party router B.AI has already moved the model off its free tier to discounted peak/off-peak pricing — a signal that serving economics shifted, but one router's pricing decision is not a benchmark.

Why builders are watching

A 1M-context MIT model with 16B active decode parameters is, if the efficiency claims hold, cheap to serve at scale — and the API price table already reflects part of that. For teams running our DeepSeek API guides, the actionable date is September 14, when V4-Pro traffic migrates and bills drop to Flash rates whether you opted in or not. The benchmark argument will settle itself as independent evaluators publish over the next week; the license and the weights, the parts that cannot be walked back, are already settled.