DeepSeek-V4-Flash Specs Leaked: 284B Parameters, FP4 Mixed Precision, Native 1M Context
vllm_project · x · 2026-08-01
The vLLM Recipes page has unexpectedly leaked detailed technical documentation for the DeepSeek-V4-Flash model. As a new member of the V4 preview family, the model has a total parameter count of 284B with only 13B active parameters.
Core Architectural Highlights:
- Hybrid Attention Stack: Combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA).
- Extreme Memory Optimization: Paired with Manifold-Constrained Hyper-Connections, it requires only 27% of V3.2's per-token inference FLOPs and 10% of its KV Cache at a 1 million token context length.
- Mixed Precision Storage: MoE expert weights are stored in FP4, while remaining parameters (attention, router, etc.) use FP8.
- Inference Acceleration: Supports Multi-Token Prediction (MTP) and DSpark speculative decoding.
Deployment & Variants:
Pre-trained on 32T+ tokens, the model utilizes a two-stage post-training pipeline. Four checkpoint variants are provided, including a native FP4+FP8 mixed weights version and an NVFP4 version re-quantized by NVIDIA, optimized specifically for Blackwell GPUs.
Related event: DeepSeek-V4-Flash Architecture Leaked with Million-Token Context(3 posts)→
More from Infra
- NVIDIA quietly delists the entry-level RTX 5050 from its marketplace, following the RTX 5090 — chemist_slime · 2026-10-16
- Grok claims 816B MoE model runs locally on dual M5 Ultra Mac Studios via MLX — grok · 2026-10-16
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16