Overview
DeepSeek V4.1 Flash became generally available on 10 September 2026, closing the two-day beta that ran under the id `deepseek-v4.1-flash-expires-on-0910`. It answers to the plain model name `deepseek-flash` on the DeepSeek API, and DeepSeek describes it as the smallest model in a new architecture family, with native multimodal visual understanding built into the base design rather than added as an extension. The weights are published on Hugging Face under an MIT license alongside a technical report.
The architecture is a causal encoder-decoder: a 20-layer encoder feeding a 20-layer decoder over a Mixture-of-Experts stack with 384 routed experts plus one shared expert per MoE layer, six routed experts active per token. That gives 552B backbone parameters and 196B additional Engram parameters, while only 8B fire during prefill and 16B during decode. Compressed Sparse Attention 2 — an FP4 KV cache with hierarchical sparse indexing — holds the global cache to roughly 890 bytes per token, about a quarter of DeepSeek V4 Flash and around 437 times smaller than DeepSeek V1, which is what makes the 1M-token window affordable to serve.
The release also reshapes DeepSeek's lineup. The V4 Flash and V4 Flash Vision Exp names are retired, with the legacy identifiers temporarily routed to V4.1 Flash, and DeepSeek is phasing out V4 Pro: after 14 September 2026 V4 Pro requests are served by V4.1 Flash at V4.1 Flash pricing. On DeepSeek's own published table at maximum reasoning effort, V4.1 Flash reaches 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1, ahead of the Opus 5.0 and GPT-5.6 Sol columns it is set against, plus 90.9 on GPQA Diamond, a 3471 Codeforces rating and 36.8 on HLE (39.1 on the pure-text subset).
On 17 October 2026 DeepSeek upgraded the model in place to the DeepSeek-V4.1-Flash-1017 checkpoint, which its model card calls a minor update. The API name `deepseek-flash` and the pricing stayed the same, the web and app models were updated too, and the 1017 weights were published on Hugging Face under the same MIT license. On DeepSeek's own table the checkpoint moves DeepSWE v1.1 from 74.2 to 77.0, Terminal-Bench 4.0 from 31.2 to 42.3 and HLE without tools from 36.8 to 42.6, and DeepSeek says it also finishes tasks faster with shorter outputs. It is tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, with DeepSeek Harness v0.2.1-rc.1 or later recommended.
| Released | 2026-09-10 |
|---|---|
| License | MIT |
| Weights | Open weights |
| Parameters | 552B backbone + 196B Engram · 8B active on prefill, 16B on decode |
| Context | 1M tokens |
| Architecture | Causal encoder-decoder Mixture-of-Experts (20-layer encoder, 20-layer decoder; 384 routed experts + 1 shared, 6 routed active per token) |
| Modalities | Text, Vision |
| Status | Generally available on the DeepSeek API Platform under the model name `deepseek-flash`, which has served the DeepSeek-V4.1-Flash-1017 checkpoint since 17 October 2026; open weights for both checkpoints are on Hugging Face. |
Benchmarks
DeepSeek's published comparison for the 1017 checkpoint
| Benchmark | DS-V4.1-Flash-1017 | DS-V4.1-Flash | Kimi K3 | GLM-5.3 | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|---|---|
| HLE (no tools) | 42.6% | 36.8% | 43.5% | 42% | 44.5% | 56.3% |
| Terminal-Bench 4.0 | 42.3% | 31.2% | 12.6% | 37.9% | 39.9% | 51.8% |
| DeepSWE v1.1 | 77% | 74.2% | 67.5% | 66.9% | 73% | 74% |
| ProgramBench | 29.1% | 20.3% | 17.5% | 19% | 23% | 37% |
| NL2Repo-Bench | 68.2% | 65.4% | 58% | 58% | 56.8% | 75.3% |
| Automation-Bench v1.0.6 | 60.5% | 54.8% | 46.7% | 48.8% | 45.8% | 50.3% |
| Agents' Last Exam | 35.4% | 31.8% | 27.6% | 28.5% | 26.7% | 28.6% |
| Apex Bench v1.1 | 72.1% | 64.6% | 47.8% | 56.6% | 54.3% | 65.8% |
| Chartography (with tools) | 83% | 78.9% | 68.1% | — | 79.9% | 84% |
| BabyVision (with tools) | 92.5% | 89.6% | 85.7% | — | 88.9% | 94.1% |
| ZeroBench-main (with tools) | 51% | 49% | 41% | — | 53% | 52% |
This model's scores
- Terminal-Bench 2.190.6%
- DeepSWE v1.174.2%
- CyberGym88.1%
- GPQA Diamond90.9%
- HLE36.8%
- MMLU-Pro (EM)74.1%
- HumanEval (Pass@1)79.4%
- GSM8K (EM)93%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Input | $0.15 / 1M tokens |
|---|---|
| Cached input | $0.003 / 1M tokens |
| Output | $0.60 / 1M tokens |
Off-peak rates; peak rates are double ($0.30 input, $0.006 cache hit, $1.20 output). Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday.
Strengths
- MIT-licensed open weights on Hugging Face, with a published technical report on the architecture
- Leads DeepSeek's own agentic comparison table: 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1, above the Opus 5.0 and GPT-5.6 Sol columns
- 1M-token context served from an FP4 KV cache of about 890 bytes per token
- Native multimodal visual understanding in the base architecture rather than a vision extension
- Flash-tier pricing from $0.15 per million input tokens off-peak, with cache hits at $0.003 per million
- Only 8B to 16B parameters active per token despite a 552B backbone
Best for
- Agentic coding and terminal work where the published Terminal-Bench and DeepSWE scores matter
- Long-document and whole-repository tasks that need the 1M-token context
- Mixed text-and-image prompts on a model that handles vision natively
- Self-hosting an open-weight frontier model through vLLM, SGLang or Transformers
- Migrating traffic off DeepSeek V4 Pro before its 14 September 2026 retirement
How to access
| Provider | Model ID |
|---|---|
| DeepSeek Platform ↗ | deepseek-flash |
DeepSeek V4 — every version
The full lineage of the DeepSeek V4 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| DeepSeek V4.1 Flashcurrent | 2026-09-10 | 1M | MIT |
| DeepSeek-V4-Flash-Vision-Exp | 2026-08-21 | 1M | MIT |
| DeepSeek-V4-Pro | 2026-08-13 | 1M | MIT |
| DeepSeek-V4-Flash | 2026-07-31 | — | MIT |
FAQ
Is DeepSeek V4.1 Flash open source?
The DeepSeek V4.1 Flash weights are published on Hugging Face under an MIT license, together with a technical report on the architecture. Running it locally is still a serious undertaking — the backbone is 552B parameters plus 196B Engram parameters, even though only 8B to 16B are active per token. The model card documents Transformers, vLLM, SGLang and Docker deployment paths.
What happens to DeepSeek V4 Pro?
DeepSeek is retiring V4 Pro. Its changelog states that after 14 September 2026, DeepSeek V4 Pro requests are routed to DeepSeek V4.1 Flash and billed at V4.1 Flash pricing. The V4 Flash and V4 Flash Vision Exp names are retired as well, with legacy identifiers temporarily pointed 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 at maximum reasoning effort, DeepSeek V4.1 Flash scores 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 reported numbers.
Why is the DeepSeek V4.1 Flash KV cache so small?
DeepSeek V4.1 Flash uses Compressed Sparse Attention 2, combining an FP4 KV cache with hierarchical sparse indexing, which holds the global cache to about 890 bytes per token. That is roughly a quarter of DeepSeek V4 Flash and about 437 times smaller than DeepSeek V1 — the compression that makes a 1M-token context practical to serve at Flash prices.
What does DeepSeek V4.1 Flash cost?
DeepSeek V4.1 Flash bills at $0.15 per million input tokens on a cache miss and $0.60 per million output tokens during off-peak hours, doubling to $0.30 and $1.20 at peak. Cache hits cost $0.003 per million off-peak. Peak hours run 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday.
What changed in DeepSeek-V4.1-Flash-1017?
It is a minor update released on 17 October 2026 that replaced the model behind `deepseek-flash` with no change to the API name or price. DeepSeek reports higher scores across its table (for example DeepSWE v1.1 77.0 vs 74.2 and Terminal-Bench 4.0 42.3 vs 31.2), plus faster task completion and shorter outputs. The weights are on Hugging Face under the MIT license.