Model · DeepSeek · Text
DeepSeek V4.1 Flash
DeepSeek's current flagship and its only actively developed API model — a natively multimodal MoE that DeepSeek says beats V4 Pro, under MIT. The 2026-10-17 update (DeepSeek-V4.1-Flash-1017) sharpened coding and agentic work and now serves the deepseek-flash API ID on web, app and API.
- Modality
- Text
- License
- MIT (Open)
- Parameter size
- 552B backbone (~763B total incl. Engram memory and vision encoder); 8B active in prefill, 16B in decode
- Context window
- 1,048,576 tokens 1M-token context. DeepSeek's pricing page lists 384k max output; the model card recommends max_tokens of 256k or more.
- Released
- September 10, 2026
- Last verified
- October 28, 2026
- Runs locally
- Yes
- Also handles
- vision, reasoning, code
What you need to run it
- Mac
- Yes, but slow. Any Apple Silicon Mac (M1 or newer) with at least 512 GB of memory. It runs, but much slower than on an NVIDIA card.
- Windows / Linux PC
- Server only. Needs a server with several graphics cards, about 1152 GB of video memory in total — not a desktop PC.
- No graphics card
- Not practical. Too slow without a graphics card or Apple Silicon.
Too big for almost any desktop: it only runs on a top-spec 512 GB Mac Studio (Ultra chip) using a heavily compressed community build, otherwise it needs a multi-GPU server.
Numbers are for: 2-bit community MLX build of the 763B model (about 313 GB); the maker's own weights are about 765 GB. Our estimate from the model’s size — the maker publishes no figure.
Strengths
- MIT licence on the weights — no commercial restrictions
- DeepSeek reports it ahead of its own V4 Pro flagship, at Flash prices ($0.15 in / $0.60 out per MTok at off-peak rates, double at peak)
- Natively multimodal — image and text input via a built-in DeepSeek-ViT encoder
- Very small active parameter count (8B prefill / 16B decode) for its capability level
- The 1017 update cut output length and task time while raising scores (HLE without tools 42.6 from 36.8; Terminal-Bench 4.0 42.3 from 31.2)
Weaknesses
- ~763B total parameters — needs multi-GPU serving, not a single-machine model
- Peak/off-peak pricing: peak hours (weekday UTC mornings) cost twice the off-peak rate
- New causal encoder-decoder architecture — check inference-engine support before self-hosting
Try it
| Where | Type | Notes |
|---|---|---|
| Hugging Face | weights | MIT licence — current 1017 update |
| DeepSeek Platform | hosted-api | API key required; model ID deepseek-flash |
| DeepSeek Chat | product | Web and app |
Used in solutions
Version history
- DeepSeek V4.1 Flash Sep 2026 Current
- DeepSeek V4 Pro Aug 2026
- DeepSeek V4 Flash Jul 2026 Deprecated
- DeepSeek V3 Dec 2024 Deprecated
Official sources
- Model card (1017) model
- Model card (original) model
- Release news announcement
- 1017 update news announcement
- Pricing docs
Change log
- — Added hardware requirements (Mac support, memory, NVIDIA). Source: size of the Hugging Face weights and mlx-community/DeepSeek-V4.1-Flash-1017-oQ2e-mtp.
- — Initial entry. Released 2026-09-10 per DeepSeek's news index; updated 2026-10-17 as DeepSeek-V4.1-Flash-1017 (coding/agentic RL update, live on web, app and API), which this entry covers rather than a separate file. API ID deepseek-flash at $0.15 in / $0.60 out / $0.003 cache-hit per MTok off-peak (double at peak), 1M context, 384k max output. The legacy IDs deepseek-v4-flash and deepseek-v4-flash-vision-exp still work but are served by V4.1-Flash-1017 at the Flash price. Parameter count: 552B backbone per the card; Hugging Face metadata shows ~763B total.
Esc