Overview
DeepSeek V4-Flash is the efficient sibling of the V4 series, released in April 2026 alongside V4-Pro. It's a 284B-parameter mixture-of-experts model with only 13B parameters active per token and a 1-million-token context window, shipping under the permissive MIT licence with open weights on Hugging Face. Where V4-Pro (1.6T total / 49B active) is the flagship, V4-Flash targets high-volume, cost-sensitive deployment while keeping most of the V4 capability.
V4 brings a new hybrid attention architecture (Compressed Sparse + Heavily Compressed Attention) that, in the 1M-token setting, cuts single-token inference FLOPs to roughly 27% and KV cache to 10% versus V3.2 on the Pro model, plus manifold-constrained hyper-connections and the Muon optimizer. V4-Flash supports three reasoning modes (non-think, think-high, think-max); with a larger thinking budget its Max mode approaches V4-Pro's reasoning quality, though its smaller scale naturally trails on pure knowledge and the hardest agentic tasks.
Key capabilities
- 284B MoE with only 13B active per token
- 1-million-token context window
- Hybrid attention — ~27% FLOPs / ~10% KV cache vs V3.2 (Pro)
- Open weights under MIT licence; three reasoning modes
At a glance
GPQA Diamond
88.1% (Flash Max)
LiveCodeBench
91.6% (Flash Max)
SWE Verified
79.0% (Flash Max)
Context
1M tokens
Pricing: Industry-low — open weights, self-hostable under MIT.
Pros & cons
What we like
- Outstanding value — frontier-adjacent quality at DeepSeek pricing
- Open MIT weights, fully self-hostable
- Efficient MoE keeps inference cheap
- Strong long-context performance
Trade-offs
- Trails V4-Pro and top closed models on the hardest knowledge tasks
- FP4+FP8 mixed precision needs compatible tooling
The verdict
DeepSeek V4-Flash is the value champion of the V4 line — a 1M-context, MIT-licensed open-weight model that approaches V4-Pro reasoning at a fraction of the size and cost. For high-volume and self-hosted workloads, it's one of the strongest open models available today.
Best for: Cost-sensitive, high-volume and self-hosted open-weight workloads.
Frequently asked questions
What's the difference between V4-Flash and V4-Pro?
V4-Pro is the flagship (1.6T total / 49B active); V4-Flash is the efficient variant (284B total / 13B active) for cheaper, higher-volume deployment. Both share the V4 architecture, 1M context and three reasoning modes — V4-Flash trails on pure knowledge and the hardest agentic tasks but approaches Pro quality with a larger thinking budget.
Can I self-host DeepSeek V4-Flash?
Yes — it ships with open weights on Hugging Face under the MIT licence, so you can self-host it subject to MIT terms. For local Think Max mode, DeepSeek recommends a context window of at least 384K tokens.
