DeepSeek-V4.1-Flash: 552B, 8B active, MIT, cheaper on the API — and V4-Flash is retired

DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026: open weights on Hugging Face under the MIT licence, and live on the DeepSeek API the same day as `deepseek-flash`. It is the smallest model in a new architecture family — 552B backbone parameters, 8B active for input and 16B for output, native image input, 1M-token context — and it is the most-downloaded model created on Hugging Face in our window by a wide margin: 288,414 downloads by September 15. The same day, DeepSeek retired V4-Flash and V4-Flash-Vision-Exp on its API and cut the Flash price.

Verdict: DeepSeek's new architecture lands as a 552B MIT model with 8B active, a quarter of V4-Flash's KV cache, and a lower API price — still big-iron for local

The take

The facts from the model card and config. `deepseek-ai/DeepSeek-V4.1-Flash`, created 2026-09-10, `license: mit` (the LICENSE file is the standard MIT text). Architecture: a 40-layer "Causal Encoder-Decoder" — a 20-layer causal encoder then a 20-layer decoder whose global KV cache is projected from the encoder's final hidden states — with 384 routed experts plus 1 shared, 6 per token, and a separate 196B-parameter "Engram" memory table read sparsely by token lookup. The safetensors total is 763B because it counts that table. DeepSeek's headline is the cache: roughly a quarter of V4-Flash's KV cache per token (DeepSeek's API notice: "1/4 the HBM, 1/8 the SSD storage"). Pre-trained from scratch on 45T multimodal tokens.

DeepSeek's benchmark claims, labelled as such — every row is vendor-run, at maximum reasoning effort, mostly on DeepSeek's own harness. On its table V4.1-Flash posts DeepSWE v1.1 74.2 (V4-Pro 62.7, Opus 5 74.0), Terminal-Bench 2.1 90.6, and Codeforces 3471, while trailing Opus 5 clearly on Terminal-Bench 3.0 and 4.0 (30.0 and 31.2 against 43.3 and 51.8) and on HLE (36.8 against 56.3). Credit where due: the card prints the rows it loses, and ships the steps to reproduce its DeepSWE result. The API notice says tests "by multiple parties" put it ahead of V4-Pro; we have not seen those tests.

The price change is the part that touches our numbers. From DeepSeek's rate card (read September 15): `deepseek-flash` costs $0.15 in (cache miss) / $0.60 out per 1M off-peak and $0.30 / $1.20 at peak, down from V4-Flash's $0.22 / $0.66 off-peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, and the card now says "Monday through Friday", so weekends are off-peak all day. The legacy names `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` still work but are served by V4.1-Flash at the new rate. Our DeepSeek Flash calculator row moves from a $0.44 to a $0.375 blended per-million rate. DeepSeek first said V4-Pro requests would route to V4.1-Flash from September 14; the rate card now says V4-Pro stays on the API "after September 14, 2026, with the billing method remaining unchanged", so that row does not move.

For local buyers: not yet, with one honest exception to watch. At 552B plus a 196B table this is past every box on our hardware list. The community build that matters is `antirez/deepseek-v4.1-flash-gguf` (207K downloads), made for the DwarfStar Metal runtime rather than llama.cpp: its card lists a 340.6 GiB Q2 file, of which 188.8 GiB is Engram data read from disk, and says the Q2 runs on "a single 128 GB Mac" with SSD streaming. That is the runtime author's claim, with no speed figure we can verify, so we do not call it a local pick. There is no Ollama local tag. Our call: no model page, no planner change; `/models/deepseek-v4-flash/` now records the successor.

Where this fits

Models: DeepSeek V4-Flash · DeepSeek V4-Pro · Kimi K3 (2.8T-A50B)

Hardware: Mac Studio M3 Ultra 96 GB · M5 Max MacBook Pro 64 GB

Sources

Next step

Try this in the planner→