LLM
GA
Open source
DeepSeek V4.1 Flash
Cheap high-volume agentic coding
Release2026-09-10
Context—
Max output—
Parameters552B
Local / VRAMLocal
LicenseOpen source
Overview
DeepSeek V4.1 Flash is a model in the LLM category from DeepSeek. It was released in 2026-09-10. The vendor is based in China. Best for: Cheap high-volume agentic coding.
Notes: The V4.1 Flash-1017 checkpoint (2026-10-17) updated the weights under the same API name and price. V4.1 Pro is announced but undated. Context window not verified (V4 line: 1M); parameter count from one source.
Hardware: About 510 GB on disk and ~614 GB of GPU memory (full 8-GPU node).
Strengths
- Ahead of V4 Pro on DeepSeek agentic benchmarks (vendor)
- Native image input
- MIT license and very low price
Limits
- Needs a full 8-GPU node to self-host
- Behind V4 Pro on knowledge tests (vendor)
Specifications
| Vendor | DeepSeek |
|---|---|
| Category | LLM |
| Release | 2026-09-10 |
| Status | GA |
| License | Open source |
| License name | MIT |
| Commercial use | Allowed |
| Parameters | 552B |
| Active parameters | 8B (input) / 16B (output) |
| Architecture | causal encoder-decoder MoE with Engram memory |
| Input modalities | Text, Image |
| Output modalities | Text |
| Prompt types | Text |
| Negative prompts | No |
| Max outputs | 1 |
| Local deployment | Yes |
| API access | Yes |
| Pricing | $0.15 / $0.60 per 1M token off-peak (peak is 2x); cache hits $0.003 |
| Confidence | Medium |
| Vendor country | China |
| Vendor type | Open lab |
Pricing & hardware
$0.15 / $0.60 per 1M token off-peak (peak is 2x); cache hits $0.003
This model can be self-hosted or run locally. API access is available.