The Open-Source AI Stack
RSS
All models

Models · deepseek-v4

DeepSeek-V4 Pro

Open DeepSeek · · MIT

First V4-generation flagship, released April 24 2026 with a 1.6T-parameter MoE activating 49B per token. Combines Compressed Sparse Attention and Heavily Compressed Attention to cut single-token inference FLOPs to 27 percent of V3.2 and KV cache to 10 percent. 1M context is now the DeepSeek default. V4-Flash (284B / 13B active) shipped alongside.

Cost

$1.74 / Mtok input
$3.48 / Mtok output

· as of 2026-05-21

source ↗

Speed

29.8 tok/sec output
1135 ms TTFT

· as of 2026-05-21

source ↗

Architecture

tokens in Embedding vocab not disclosed · deepseek tokenizer × N layers Attention (not disclosed) Position encoding not disclosed context 1,000,000 tokens MoE Router ? experts total · ? active per token Output projection tokens out
Schema-generated from data/models.yaml. Every label is auditable against the model's sources.

Specs

Architecture
moe
Total params
1.6T
Active params
49B
Context window
1.0M tokens
Attention
hybrid-csa-hca
Position encoding
unknown
Pretraining tokens
32.0T
Post-training
sft, grpo, distillation
OSI-approved
yes
Data released
no
Training code
not released

Benchmarks

Each score carries the date it was published; we never infer or interpolate missing scores.

General reasoning

MMLU-Pro 87.5 as of 2026-04-24 source ↗
GPQA-Diamond 88.8 as of 2026-05-21 source ↗

Code

HumanEval 76.8 as of 2026-04-24 source ↗
SWE-Bench Verified 80.6 as of 2026-04-24 source ↗
LiveCodeBench 93.5 as of 2026-04-24 source ↗

Available quantizations

GGUF llama.cpp's container; the common local format, k-quants from Q2 to Q8. runs on llama.cpp, Ollama
MLX Apple MLX 4/8-bit layout for Apple silicon. runs on Apple MLX
FP8 8-bit float, frequently a native release on Hopper / Blackwell GPUs. runs on vLLM, SGLang, TensorRT-LLM

Verified via the Hugging Face model tree ↗. Community quantizations change over time; the families shown are those with published weights at audit time.

Notable innovations

  • · Hybrid CSA + HCA attention
  • · Manifold-Constrained Hyper-Connections (mHC)
  • · Muon optimizer at this scale
  • · FP4 + FP8 mixed-precision MoE training
  • · 1M context as default

Lineage

First V4 flagship; new hybrid attention and FP4 training stack.

Sources