DeepSeek V4.1 Flash Launches—Reconstructed From Third-Party Trackers

Important caveat: we could not retrieve a first-hand [DeepSeek](/tools/deepseek-app) announcement. The facts below come from four independent third-party model trackers—BenchLM, All Agent, Yotta Labs, and Get Model Key—where they agree; anything they dispute or omit is left out.

What all four trackers agree on#

• Launch and date: DeepSeek V4.1 Flash went live on the official API on September 10, 2026 under the model ID deepseek-flash (all four log the date; three log the ID).
• Open weights: full weights are hosted on Hugging Face under the MIT license and can be self-hosted.
• Specs: a 552B-parameter mixture-of-experts backbone with a new causal encoder-decoder architecture, activating ~8B parameters per token on prefill and ~16B on decode; a 1M-token context window, 384K max output, and native image input.
• One benchmark: several trackers list Terminal-Bench 2.1 at 90.6—but all note this is vendor-reported, with no independent replication at launch.

Pricing and one reversal#

USD and CNY pricing figures describe the same official tiers: roughly $0.15 input / $0.60 output per million tokens off-peak (about ¥1 / ¥4), doubling at peak, with cache-hit input around $0.003. Note the "V4 Pro routing" reversal: per Yotta Labs and All Agent, DeepSeek's launch notice said that from September 14 deepseek-v4-pro requests would route to V4.1 Flash at Flash prices, then walked that back before it took effect, keeping V4 Pro live with unchanged billing. We have only tracker accounts of this, not the original notice.

How we read it#

As analysis, not a verified report, two things stand out. First, a model the media call a flagship replacement can currently only be pieced together from third-party trackers—which says a lot about today's open-model information ecosystem. Second, the vendor claim of beating V4 Pro on "performance, cost, speed, and task-completion time" has no independent corroboration across the four; Yotta Labs relays that Artificial Analysis later measured it at 40 on its Intelligence Index (above V4 Flash's 35, just under GLM 5.3 Flash's 42). Until we can reach DeepSeek's own material, treat this as analysis, not a confirmed announcement.