DeepSeek-V4.1-Flash: a 552B multimodal MoE under MIT, built to run the decoder cheap

What happened

DeepSeek published DeepSeek-V4.1-Flash to Hugging Face on 10 September 2026: a 552-billion-parameter multimodal Mixture-of-Experts model that takes images and text in and produces text out, under an MIT licence. At collection the hub showed 840 likes against six downloads — a fresh release with early traction but no deployment trail yet.

Context

The V4-Flash line already put sparse-attention MoE models in the open-weight space. V4.1-Flash restructures the core: a 40-layer Causal Encoder-Decoder Transformer split 20 layers to the encoder and 20 to the decoder, a 196-billion-parameter Engram module for conditional memory, and a vision encoder trained from scratch on a 45-trillion-token multimodal corpus. Post-training is a straightforward SFT → RL → on-policy distillation sequence with no algorithmic modifications. Sparse attention was trained at 64K sequence length, then the context was extended to one million tokens over 34T additional tokens. The MIT licence in the hub tags is the detail that matters most for anyone planning to build on it rather than just benchmark it.

How it works

The encoder does the heavy representational work; the decoder reads projected hidden states. During prefill, 8B parameters are active per token; during decode, 16B. Each MoE layer carries one shared expert plus 384 routed experts, with six activated per token. The 196B Engram component is sparsely accessed by token-based lookup rather than dense attention, so it does not add a per-token compute cost the way a dense layer would.

Sparse attention is handled by Compressed Sparse Attention 2, which assigns each layer one of three static modes: Full, Reindex, or Reuse. SWA Bounded Replay cuts the persistent KV cache to roughly one-eighth of V4-Flash's footprint. The global KV cache sits at 890 bytes per token, approximately one-quarter of V4-Flash. KV storage uses FP4 in E2M1 format with one E4M3 scale per 16 channels.

Image input goes through a DeepSeek-ViT encoder using 2D-RoPE and 3×3 pixel-unshuffle downsampling, then a two-layer MLP projector maps visual patches into the text embedding space. The model exposes a reasoning-effort setting as an integer from 1 to 100. Inference runs through the transformers library; the hub tags list 8-bit and fp8 checkpoint formats.

Our read

The CED split is the detail the headline buries. In a standard decoder-only MoE, every token re-derives its attention context from the full sequence. Here, the 20-layer encoder does the representational work and the 20-layer decoder completes generation from projected hidden states. That is why active parameters are 8B at prefill and 16B at decode: the decoder is not re-computing the sequence, it is reading a compressed summary. The 552B count is a memory figure, not a per-token FLOPs figure, and treating it as one is the most common error you will see in early reviews.

That distinction decides whether this is "impossible locally" or "possible on a two-GPU node." The 890-byte KV cache at one-million-token context works out to roughly 890 MB for a full pass — small enough for a single high-end accelerator. The binding constraint is the weight file itself, and the hub tags list 8-bit and fp8 formats without stating the actual .safetensors sizes. That number, not the parameter count, is what you need before committing hardware.

The reasoning-effort integer (1–100) is the quiet feature. It is a continuous cost/quality dial: set it to 15 for generating fifty scene-idea variants, crank it to 90 for the final script polish. Most open-weight models offer a binary "think or don't think" toggle via system prompt; a continuous knob is a cleaner interface for a pipeline that varies effort per call.

What this changes

For a ComfyUI-based studio, this is a script and direction generator, not a frame generator. The image-text-to-text task means you feed it reference frames or storyboard panels and get back shot descriptions, continuity notes, or prompt text for your actual video model. It does not produce pixels.

Practically: expect to call it through a transformers-based Python node or an API endpoint — no ComfyUI-native node is implied. The 1M context window is headroom you will not need for a typical five-to-fifteen-second clip; it becomes relevant only when you feed a full multi-episode storyboard in a single prompt. Verify the actual checkpoint file sizes before allocating VRAM, since the hub does not state them. If your current LLM scripting step is already producing usable output, nothing in this release changes what you do on Monday.

License

MIT. Commercial use, modification, and redistribution are permitted without a revenue ceiling, entity-size restriction, or attribution condition beyond the licence text itself. For a studio shipping a pipeline that ingests this model's output, there is no licence-review step to schedule.

Key takeaways

• DeepSeek-V4.1-Flash is a 552B multimodal MoE (image + text in, text out) under MIT licence, published 10 September 2026 with a linked technical report.
• The 20+20 Causal Encoder-Decoder split means 8B/16B active parameters per token, not 552B — the parameter count is a memory figure, not a per-token compute figure.
• The 890-byte KV cache at 1M context (~890 MB) is small; the binding constraint for local inference is the weight file size, which the hub does not yet state.
• The reasoning-effort integer (1–100) gives a continuous cost/quality dial absent in most open-weight models.
• This is a text-output model: it generates scripts, directions, and prompts from visual input. It does not produce video frames.

Sources

1. deepseek-ai/DeepSeek-V4.1-Flash — image-text-to-text on Hugging Face — tier 1
2. deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face — tier 3