DeepSeek V4 Flash
DeepSeek V4 Flash is a Mixture-of-Experts text-generation model in the DeepSeek-V4 series from deepseek-ai, released under the MIT license. According to its official model card, it has 284B parameters with 13B activated and supports a context length of one million tokens. It uses a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), Manifold-Constrained Hyper-Connections (mHC), and the Muon optimizer. The card states it was pre-trained on more than 32T tokens and underwent a two-stage post-training pipeline. The card also refers to a Flash-Max maximum reasoning effort mode and reports that Flash-Max achieves comparable reasoning performance to the Pro version when given a larger thinking budget. The repository contains safetensors weights and is tagged for text-generation with Transformers as the library name.
At a glance
- Parameters
- 290.94B
- Architecture
- deepseek_v4
- Context length
- 1,048,576
- License
- mit
- Software
- Publisher PyTorch reference
Will it run on your machine?
Save your machine to see a personalized rating and its reasoning.
Add my hardwareBest for
- Text generation
- Long-context tasks up to one million tokens as stated on the model card
- Reasoning-oriented generation via the referenced Flash-Max maximum reasoning effort mode
- Load into Transformers for inference (per library_name on the model card)
- Use cases associated with a Mixture-of-Experts language model as described
Tradeoffs
- Only model-card facts are summarized; no independent testing was performed
- No numerical hardware requirements, memory needs, or compatibility details are provided in the source facts
- The card provides FLOP/KV-cache comparisons only for DeepSeek-V4-Pro, not for Flash
- The supplied model card text is truncated, so the full documentation and evaluation details are not available
- The card's performance or benchmark claims are its own statements and are not independently verified here
Ways to run it
Publisher PyTorch reference · Linux / rented GPU
Repository-specific command in publisher documentation; confirm dependencies and hardware in the source.
torchrun --nproc-per-node ${MP} generate.py --ckpt-path ${SAVE_PATH} --config ${CONFIG} --interactiveExplore its uses
Keep exploring
Bonsai 2 · 27B
A compressed 27B-class reasoning model with publisher-provided GGUF packs and a dedicated llama.cpp fork for CUDA, Metal, and CPU.
DeepSeek V4.1 Flash
A multimodal reasoning model with a compressed key-value cache, published as open weights.
Qwen3.8-Flash-Next
An experimental open-weight multimodal model with sparse attention and 262K native context.