Model guide · September 2026
DeepSeek V4.1 Flash, fully specified
DeepSeek's 552B mixture-of-experts shipped 1M context and native vision at a small fraction of what frontier models charge — and buried a peak-hour multiplier in the pricing page that quietly doubles most cost estimates.
The short version
Published rates are $0.15 input and $0.60 output per million tokens, but those are off-peak: DeepSeek doubles both to $0.30 / $1.20 during weekday peak windows. Cache hits cost $0.003 with no cache-write fee, there is no batch API, and the weights are MIT-licensed at roughly 510 GB in FP8. DeepSeek's own evals claim it beats the pricier V4 Pro on agentic work.
Full specification
Every figure below is DeepSeek's own published number, read from the same model registry this site prices everything with.
| Specification | DeepSeek V4.1 Flash |
|---|---|
| Release date | 2026-09-10 |
| API model name | deepseek-flash |
| Input (off-peak) | $0.15 |
| Output (off-peak) | $0.60 |
| Peak-hour billing | 2x off-peak, Mon–Fri 01:00–04:00 and 06:00–10:00 UTC |
| Cache hit | $0.003 |
| Cache write | No charge — prefix caching is automatic |
| Batch API | Not offered |
| Architecture | 552B MoE — 8B active on prefill, 16B on decode |
| Weights | MIT open weights (about 510 GB at FP8) |
| Context window | 1M |
| Max output | 384K |
| KV-cache footprint | 890 B/token |
| Data retention | None published |
Verified against DeepSeek's published pricing page on 2026-09-19. All rates are USD per million tokens.
Off-peak against peak pricing
On weekdays, 01:00–04:00 and 06:00–10:00 UTC — business hours in Beijing — every rate doubles, the cache-hit price included. Outside those windows the published rates apply.
| Rate | Off-peak | Peak hours |
|---|---|---|
| Input / 1M tokens | $0.15 | $0.30 |
| Output / 1M tokens | $0.60 | $1.20 |
| Cache hit / 1M tokens | $0.003 | $0.006 |
Claude Fable 5.1 lists at 67x this model's input rate and 83x its output rate — the anchor most people searching this model are really after.
Both columns are vendor-published rates verified 2026-09-19; the comparison figure was verified 2026-09-14. DeepSeek publishes no batch discount, so the peak multiplier lands on whatever share of your spend falls inside those hours.
Self-hosting it
What DeepSeek has actually published, and where the honest answer stops.
The download is the easy part
The weights are MIT-licensed at roughly 510 GB in FP8, so the licence puts no restriction on what you serve with them. Size is the release's real constraint, not the terms.
A mostly-inactive model is still a huge model
Only 8B of 552B parameters are active per prefill token and 16B during decode, but the whole weight set has to sit in high-bandwidth memory regardless. Sparse activation cuts the compute per token; it does not cut the VRAM bill, which is what makes this a multi-GPU deployment rather than a workstation one.
We have not verified the serving recipe
Hardware counts, quantisation choices and throughput claims for this model are not something we have reproduced ourselves, so they are not on this page. Single-node numbers circulating in GPU-cloud marketing deserve the same suspicion. A step-by-step guide lands here once we have run it.
Working with the model
Three details that bite people the first week.
Two names, one model
The release is marketed as V4.1 Flash; the model name you pass to the API is deepseek-flash. Requests sent to the retired V4 Flash name now route here, which is why an end-of-life model's identifier still answers.
Vision is on the same price
Image input is native rather than a separate multimodal endpoint, and DeepSeek publishes no surcharge for it — the input and output rates above cover both modalities.
The retention difference is the one that matters
DeepSeek publishes no mandatory retention window for this model. Claude Fable 5.1 carries a 30-day one. If you are choosing on compliance rather than price, that constraint moves before any of the numbers above do.
What to watch next
Two things move here. The peak-hour window is the single largest error in most V4.1 Flash budgets, so re-check it whenever your traffic shifts timezone. And the V4 Pro retirement that started this cycle was set for 2026-09-14 and then cancelled on 2026-09-11 — if you re-architected around the announced removal, verify the current routing before you rebuild anything.
Pricing it against the alternative
A per-token sheet says nothing about your monthly bill. Both of these turn the rates above into a number you can argue with.