DeepSeek V4.1 Flash
The current fast DeepSeek. Index 39 at max, about 209 tokens per second, and one of the cheapest measured tasks among models that are actually capable.
No High score is published. Showing Max.
Every published effort
| Setting | Index | Per task | Tokens / sec |
|---|---|---|---|
| Max | 39 | $0.27 | 209 |
| Published | 25 | $0.15 | 202 |
What this workload costs
List price for one call. Reasoning tokens count as output. A cache hit uses the lab’s cache rate when one is published.
Specs
- API id
- deepseek-flash
- Released
- 10 Sep 2026
- Context
- 1M
- Max output
- 384k
- Input / 1M
- $0.30
- Output / 1M
- $1.20
- Cache read / 1M
- $0.0060
- Knowledge
- —
- Inputs
- Text, Image
- Weights
- Open weights. The V4 Flash line is reported as MIT. Confirm the card for this checkpoint.
Official API, peak cache-miss rate. Off-peak is half. Peak hours have been described as weekday 01:00–04:00 and 06:00–10:00 UTC. Cache hits are a fraction of a cent. Third-party hosts often undercut the official API because the weights are open. Max output figure is from the V4 pricing write-up.
Use it when
Fast open-weight calls with vision, a 1M window, and a cache so cheap that long prompts stop dominating the bill. The non-reasoning score on the leaderboard is 25. Use the reasoning setting if you want the 39.
Leave it when
Frontier agent work where the ten-point gap to Sol or Grok shows up. Also, budgeting only the off-peak rate if you do not control the clock.