DeepSeek V4
DeepSeek V4 Release & Features | 1.6T-Param MoE, 1M Context, Open-Source MIT
DeepSeek V4 is here: 1.6T-parameter MoE (49B active), 1M-token context, CSA+HCA hybrid attention, SWE-bench 80.6% | Released and open-sourced April 24, 2026
DeepSeek V4 was officially released and open-sourced under the MIT license on April 24, 2026, with weights on Hugging Face. It ships in two versions: V4-Pro (1.6T total / 49B active) for high-end reasoning and agentic coding, and V4-Flash (284B / 13B) for faster, lower-cost use. Both offer a 1M-token context window powered by a hybrid attention architecture (CSA + HCA) that cuts per-token compute to ~27% of V3.2 and KV-cache memory to ~10% at 1M context. V4-Pro scores 80.6% on SWE-bench Verified — the highest among open models, tied with Gemini 3.1 Pro (80.6%) and ahead of GPT-5.4 (77.2%). API pricing is $0.435/$0.87 per M tokens (Pro) and $0.14/$0.28 (Flash), roughly 5-30x cheaper than closed frontier models.
📅 Release Timeline
DeepSeek-V3 Released
671B params, 37B active, MoE architecture
MODEL1 Code Appears
MODEL1 identifier found in GitHub FlashMLA repo
V4 Released & Open-Sourced
Launched April 24, 2026 (MIT). V4-Pro 1.6T/49B, V4-Flash 284B/13B, 1M context
Enterprise Version Live
Atlas Cloud syncs V4 enterprise service on release
🚀 Core Features
From the official April 24, 2026 release
Two Versions: Pro & Flash
DeepSeek V4 ships in two open-source versions: V4-Pro for high-end reasoning and agentic coding, and V4-Flash for faster, lower-cost workloads. Both focus on text, code and reasoning.
- • V4-Pro: 1.6T total / 49B active parameters
- • V4-Flash: 284B total / 13B active parameters
- • Both support a 1M-token context window
- • Focused on text, code and reasoning
1.6 Trillion Parameter MoE
V4-Pro features 1.6 trillion total parameters with only 49B active per token via Mixture-of-Experts. This delivers frontier performance at a fraction of closed-model cost.
- • V4-Pro: 1.6T total, 49B active per token
- • Mixture-of-Experts (MoE) architecture
- • API pricing: $0.435/$0.87 per M tokens (Pro)
- • Open-source under MIT license
Million-Token Context
Both V4-Pro and V4-Flash support a 1M-token context window (max output ~384K), enough to process entire books, large codebases or ultra-long documents.
- • 1M-token context by default on both versions
- • Max output around 384K tokens
- • Process entire books (~500K words)
- • Analyze complete large project codebases
CSA + HCA Hybrid Attention
V4 combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) on top of its MoE design, making a 1M-token context window practical and cheap.
- • Hybrid of CSA and HCA attention
- • ~27% of V3.2 per-token compute at 1M context
- • ~10% of V3.2 KV-cache memory at 1M context
- • Enables low-cost ultra-long context
Ultra-Efficient 1M Context
Thanks to the CSA+HCA hybrid attention, V4 serves a full 1M-token context at a fraction of prior compute and memory cost — the core efficiency story of the release.
- • Per-token compute ~27% of V3.2 at 1M context
- • KV-cache memory ~10% of V3.2 at 1M context
- • Makes million-token workloads affordable
- • Strong throughput for agentic, long-context tasks
Strong Reasoning & Math
V4 posts strong reasoning and math results: GPQA Diamond 90.1%, MMLU-Pro 87.5%, GSM8K 92.6%, and Terminal-Bench 2.0 67.9% for agentic tasks.
- • GPQA Diamond: 90.1%
- • MMLU-Pro: 87.5%
- • GSM8K: 92.6%
- • Terminal-Bench 2.0: 67.9%
Far Cheaper Than Closed Models
DeepSeek V4-Pro API pricing is $0.435/M input and $0.87/M output (after a 75% cut); V4-Flash is $0.14/$0.28. Open-source means free self-hosting.
- • V4-Pro: $0.435 input / $0.87 output per M tokens
- • V4-Flash: $0.14 input / $0.28 output per M tokens
- • Roughly 5-30x cheaper than closed frontier models
- • Open-source (MIT): free to self-host
Leads Open Models in Coding
V4-Pro scores 80.6% on SWE-bench Verified — the highest among open models, tied with Gemini 3.1 Pro (80.6%) and ahead of GPT-5.4 (77.2%).
- • SWE-bench Verified: 80.6% (highest open model)
- • LiveCodeBench Pass@1: 93.5
- • Codeforces rating: 3206
- • Strong multi-language and repo-level coding
🔬 Technical Deep Analysis
Inside the V4 hybrid attention architecture
Architecture Innovation
- ✓ MoE backbone with two sizes (1.6T/49B Pro, 284B/13B Flash)
- ✓ Hybrid attention combining CSA and HCA
- ✓ 1M-token context window by default
- ✓ Open-sourced under the MIT license
Efficiency at 1M Context
- ✓ Per-token compute ~27% of V3.2
- ✓ KV-cache memory ~10% of V3.2
- ✓ CSA compresses sparse long-range attention
- ✓ HCA heavily compresses retained context
Performance & Cost
- ✓ SWE-bench Verified 80.6% (highest open model)
- ✓ LiveCodeBench Pass@1 93.5, Codeforces 3206
- ✓ V4-Pro pricing $0.435/$0.87 per M tokens
- ✓ V4-Flash pricing $0.14/$0.28 per M tokens
📊 V3 vs V4 Comparison
Main upgrade points overview
🏆 V4 vs Frontier Models
How DeepSeek V4 stacks up against GPT-5.4, Claude 4.6, and Gemini 3.1 Pro
📎 Information Sources
The following is based on DeepSeek's official release (2026-04-24)
Official Release
- • Released and open-sourced (MIT) on April 24, 2026, weights on Hugging Face
- • V4-Pro 1.6T/49B and V4-Flash 284B/13B, both 1M-token context
- • CSA + HCA hybrid attention: ~27% compute, ~10% KV-cache memory vs V3.2 at 1M
Official Benchmarks
- • SWE-bench Verified 80.6% (highest open model, tied with Gemini 3.1 Pro)
- • LiveCodeBench Pass@1 93.5, Codeforces 3206
- • MMLU-Pro 87.5%, GPQA Diamond 90.1%, GSM8K 92.6%, Terminal-Bench 2.0 67.9%
Official Pricing & Access
- • V4-Pro $0.435/$0.87 per M tokens, V4-Flash $0.14/$0.28 per M tokens
- • Available via chat.deepseek.com (Expert/Instant Mode), official API, Atlas Cloud
- • Legacy deepseek-chat and deepseek-reasoner retire on July 24, 2026
🎁 How to Use V4 Today
DeepSeek V4 is live on Atlas Cloud
Register Atlas Cloud Now
Register an account and get free credits
Get Your API Key
Create an API key in the console
Switch Model
Set model to 'deepseek-v4-pro' (or 'deepseek-v4-flash') in your API request
📬 Subscribe to V4 Updates
Get DeepSeek V4 news, tutorials and updates