Sulat.com
AI models
$10 off the fastest DeepSeek V4.1 Flash, Kimi K3 and GLM 5.3
$10 off the fastest DeepSeek V4.1 Flash, Kimi K3 and GLM 5.3 from Synthetic
LLM Gateway logo

Model details

DeepSeek V4 Pro (DeepSeek)

DeepSeek V4 Pro is positioned as the capable tier of the V4 family, with a Mixture-of-Experts design that keeps 49 billion of its 1.6 trillion parameters active during inference. Its architecture combines Compressed Sparse Attention and Heavily Compressed Attention, while Manifold-Constrained Hyper-Connections strengthen residual information flow. The reported use of the Muon optimizer also points to an emphasis on training stability and convergence.

The model is aimed at demanding reasoning and agentic workloads, including applications that need to work across very long inputs. Its technical report describes support for one million tokens and reports substantially lower inference compute and KV-cache requirements than DeepSeek V3.2 in that setting. The later availability update also characterizes it as improving production-agent performance, making it a practical fit for code analysis, document-intensive tasks, and multi-step workflows that benefit from sustained context.

LLM Gatewaydeepseek/deepseek-v4-prodeepseek-thinking
Open provider pageProvider docs

Quick Info

Powered by
Provider
LLM Gateway
Model key
deepseek/deepseek-v4-pro
Release date
Apr 24, 2026
Last updated
Apr 24, 2026
Knowledge cutoff
2025-05
Input modalities
Output modalities
Capabilities

Cost

Input token cost
$0.435
Output token cost
$0.87

Limits

Output tokens
393,216 tokens
Context window
1,050,000 tokens

Transparent token rates

Compare DeepSeek V4 Pro (DeepSeek) pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family
News search byUse it for free

Latest news about DeepSeek V4 Pro (DeepSeek)

LLM Gateway

CoverageLeaks

This third-party roundup tracks the state of DeepSeek V4 ahead of a confirmed general-availability date. On April 24, 2026, DeepSeek shipped a V4 Preview comprising two open-weight MIT-licensed models — V4-Pro (1.6T total, 49B active) and V4-Flash (284B total, 13B active) — both defaulting to a 1M-token context window For developer-relevant signal, the piece cites V4-Pro at 80.6% on SWE-bench Verified, within 0.2 points of Claude Opus 4.6 (80.8%), at roughly one-seventh of the output price, and frames the attention rework as targeting serving-cost reduction rather than raw capability. Reception is reported as muted relative to the R

LLM Gateway

CoverageBenchmark

This third-party benchmark review profiles DeepSeek-V4-Pro as a 1.6T-parameter MoE model with 49B activated parameters per inference, released by DeepSeek-AI on April 24, 2026 with a knowledge cutoff of May 2025. The architecture is detailed as 384 routing experts plus one shared expert, paired with CSA and HCA hybrid The article benchmarks V4-Pro horizontally against GLM 5.1 and Kimi K2.6 and vertically against earlier DeepSeek V3-series models, drawing on DataLearner's 2026 evaluation library. The 1M-token context represents a roughly 4x to 8x expansion over DeepSeek V3.2's 256K limit, enabling processing of large technical docume

Videos about DeepSeek V4 Pro (DeepSeek)

More models around DeepSeek V4 Pro (DeepSeek)