DeepSeek · 2026-09-10 · seismic
DeepSeek V4.1 Flash — a 552B open-weight rebuild with 1M context
DeepSeek V4.1 Flash is now generally available under the API name deepseek-flash, with MIT open weights on Hugging Face. It scores 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1, ahead of Opus 5.0 and GPT-5.6 Sol.
DeepSeek's smallest new-architecture model is now generally available with MIT open weights, a 1M-token context and top agentic scores.
Key specs
| Kv cache | 890 bytes per token |
|---|---|
| Terminal bench 2.1 | 90.6 |
Quick facts
| Maker | DeepSeek |
|---|---|
| API model name | deepseek-flash |
| Parameters | 552B backbone + 196B Engram; 8B active on prefill, 16B on decode |
| Context window | 1M tokens |
| License | MIT, open weights on Hugging Face |
| Price (input) | $0.15 / 1M tokens off-peak, $0.30 peak |
| Price (output) | $0.60 / 1M tokens off-peak, $1.20 peak |
Benchmarks
| DeepSeek-V4.1-Flash | 90.6% | |
|---|---|---|
| Opus 5.0 | 89.1% | |
| GPT-5.6 Sol | 88.8% | |
| DeepSeek-V4-Flash | 82.7% |
Pricing
| Input (cache miss) · Off-peak; $0.30 at peak | $0.15 / 1M tokens |
|---|---|
| Input (cache hit) · Off-peak; $0.006 at peak | $0.003 / 1M tokens |
| Output · Off-peak; $1.20 at peak | $0.60 / 1M tokens |
What is it?
General availability arrived on September 10: DeepSeek V4.1 Flash replaces the two-day beta endpoint and answers to the plain model name deepseek-flash. DeepSeek calls it the smallest model in a new architecture family, with native multimodal visual understanding rather than image support bolted on. The weights are on Hugging Face under an MIT license alongside a technical report.
How does it work?
A causal encoder-decoder design — a 20-layer encoder feeding a 20-layer decoder — sits on top of a Mixture-of-Experts stack with 384 routed experts plus one shared expert per layer, six of them active per token. That gives 552B backbone parameters and 196B Engram parameters while only 8B fire during prefill and 16B during decode. Compressed Sparse Attention 2 with an FP4 KV cache holds the whole cache to about 890 bytes per token, which is what makes the 1M-token window practical.
Why does it matter?
An MIT-licensed open model now leads DeepSeek's published agentic table against closed frontier systems: 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1, ahead of Opus 5.0 and GPT-5.6 Sol. It is also the cheap tier, at $0.15 per million input tokens off-peak. DeepSeek is retiring V4 Pro behind it — after September 14 those requests land on V4.1 Flash at V4.1 Flash prices — so the expensive tier is being folded into the cheap one.
Who is it for?
developers on the DeepSeek API and teams self-hosting open-weight models
Frequently asked questions
- Is DeepSeek V4.1 Flash open source?
- The DeepSeek V4.1 Flash weights are published on Hugging Face under an MIT license, so you can download, run and redistribute them. DeepSeek also released a technical report on the model page describing the architecture. Running it locally is a serious undertaking: the backbone is 552B parameters plus 196B Engram parameters, even though only 8B to 16B are active per token.
- What happens to DeepSeek V4 Pro?
- DeepSeek is retiring V4 Pro. The changelog says that after September 14, 2026, requests to DeepSeek V4 Pro are routed to DeepSeek V4.1 Flash and billed at V4.1 Flash pricing. The older V4 Flash and V4 Flash Vision Exp names are also retired, with legacy names temporarily pointing at V4.1 Flash so existing code keeps working.
- How does DeepSeek V4.1 Flash compare with Opus 5.0 and GPT-5.6 Sol?
- On DeepSeek's published table, DeepSeek V4.1 Flash reaches 90.6 on Terminal-Bench 2.1 against 89.1 for Opus 5.0 and 88.8 for GPT-5.6 Sol, and 74.2 on DeepSWE v1.1 against 74.0 and 73.0. On CyberGym it posts 88.1 against 84.5 for GPT-5.6 Sol. These are DeepSeek's own numbers at maximum reasoning effort.
- Why is the DeepSeek V4.1 Flash KV cache so small?
- DeepSeek V4.1 Flash uses Compressed Sparse Attention 2 with an FP4 KV cache and hierarchical sparse indexing, which brings the global cache down to about 890 bytes per token. That is roughly a quarter of DeepSeek V4 Flash and about 437 times smaller than DeepSeek V1, which is what makes a 1M-token context affordable to serve.
- How do you run DeepSeek V4.1 Flash yourself?
- The DeepSeek V4.1 Flash model card documents deployment through Transformers, vLLM, SGLang or Docker, and ships an encoding.py helper for prompt formatting. DeepSeek recommends temperature 1.0, top_p 0.95 and a max_tokens setting of at least 256K. The hosted alternative is the DeepSeek API under the model name deepseek-flash.
Try it
model="deepseek-flash" against https://api.deepseek.com