DeepSeek V4 Full Specifications: 1.6T MoE, CSA+HCA, 1M Token Context (Released 2026)
On April 24, 2026, DeepSeek officially launched and open-sourced the DeepSeek V4 preview, releasing the weights under the MIT license on Hugging Face. The pre-launch speculation has now been settled: V4 does not win on "infinite memory" or "native multimodality." Instead, it is built around agentic coding, a 1-million-token long context, and extreme inference efficiency — redefining the price-performance ratio of open-source frontier models. This article provides an authoritative breakdown of V4's full specifications based on the official release.
Dual-Version Design: Pro and Flash
DeepSeek V4 shipped in two versions at once, targeting different compute and cost scenarios:
| Version | Total Parameters | Active Parameters | Positioning |
|---|---|---|---|
| DeepSeek-V4-Pro | 1.6 trillion (1.6T) | 49B | High-end reasoning and agentic coding |
| DeepSeek-V4-Flash | 284B | 13B | Faster, lower-cost everyday workloads |
Both versions use a MoE (Mixture-of-Experts) architecture and offer a default 1-million-token (1M) context window, with a maximum output of around 384K tokens. Pro targets the strongest reasoning and coding capability, while Flash dramatically lowers latency and cost while maintaining high quality — ideal for high-concurrency, latency-sensitive applications.
When a single comparison figure is needed, V4-Pro's 1.6T total / 49B active parameters are the reference point.
Hybrid Attention Architecture: CSA + HCA
V4's real efficiency breakthrough lies in its hybrid attention architecture, which combines two attention mechanisms:
- CSA (Compressed Sparse Attention): Sparsifies long sequences, computing attention only between relevant positions to dramatically cut the compute cost of long contexts.
- HCA (Heavily Compressed Attention): Heavily compresses the KV representations, substantially reducing memory footprint.
Combined, these mechanisms allow V4 to process a 1M-token context using roughly 27% of the per-token compute of V3.2 and about 10% of its KV cache memory.
| Metric (1M context) | DeepSeek V3.2 | DeepSeek V4 |
|---|---|---|
| Per-token compute | 100% | ~27% |
| KV cache memory | 100% | ~10% |
| Context window | — | 1M tokens |
This is precisely why V4 can offer a million-token context at such a low price: not through an external "memory database," but through the structural efficiency of the attention mechanism itself.
Real Pricing: A Benchmark for Open-Source Value
V4's API pricing was already cut by 75% at launch, settling into its long-term tier:
| Model | Input Price (/M tokens) | Output Price (/M tokens) |
|---|---|---|
| DeepSeek-V4-Pro | $0.435 | $0.87 |
| DeepSeek-V4-Flash | $0.14 | $0.28 |
| GPT-5.4 | $2.50 | $15.00 |
| Claude 4.6 (Opus) | $5.00 | $25.00 |
| Gemini 3.1 Pro | $2.00 | $12.00 |
Comparing V4-Pro against closed-source frontier models, input pricing is roughly 5-12x cheaper and output pricing roughly 14-29x cheaper; V4-Flash is even lower, and real-world costs drop further with cache-hit discounts. For users who self-deploy under the MIT license, there are no API fees at all.
Benchmarks: Leading on Agentic Coding and Reasoning
The following are V4-Pro's published, real benchmark scores — no longer pre-launch "targets" or "estimates":
| Benchmark | DeepSeek V4-Pro |
|---|---|
| SWE-bench Verified | 80.6% (highest among open models, tied with Gemini 3.1 Pro) |
| LiveCodeBench Pass@1 | 93.5 |
| Codeforces Rating | 3206 |
| MMLU-Pro | 87.5% |
| GPQA Diamond | 90.1% |
| GSM8K | 92.6% |
| Terminal-Bench 2.0 | 67.9% |
The 80.6% on SWE-bench Verified is the highest score among open-source models, tied with Gemini 3.1 Pro, placing V4 firmly in the top tier for real-world software engineering tasks. The Codeforces rating of 3206, LiveCodeBench score of 93.5, and Terminal-Bench 2.0 result of 67.9% together confirm V4's strength across competitive algorithms, code generation, and terminal-agent tasks.
Open Source and Access
DeepSeek V4 is open-sourced under the MIT license, with weights released on Hugging Face. This means:
- Fully free commercial use: Enterprises can integrate V4 into products with no added restrictions.
- Free to modify and distribute: Researchers can fine-tune, distill, and build on V4.
- Local deployment: Run entirely on your own infrastructure for maximum data privacy.
The main ways to access V4:
- chat.deepseek.com: Offers Expert Mode and Instant Mode.
- Official API: Use model names such as
deepseek-v4-pro; the legacydeepseek-chatanddeepseek-reasonerwill be retired on July 24, 2026. - Third-party platforms like Atlas Cloud: atlascloud.ai is typically among the first to offer new DeepSeek models.
Summary: V4's Core Value
The release of DeepSeek V4 makes the competitive focus of open-source frontier models clear:
- Dual-version design: Pro (1.6T/49B) and Flash (284B/13B) cover everything from high-end reasoning to low-cost workloads.
- 1M-token context: Long documents, whole-repo code, and multi-turn agentic tasks are all within reach.
- CSA + HCA hybrid attention: Extreme long-context efficiency at ~27% compute and ~10% KV memory.
- Agentic coding leadership: 80.6% on SWE-bench Verified, the highest among open models.
- Ultra-low pricing + MIT open source: Value and openness combined, advancing accessible AI.
V4 no longer relies on the pre-launch packaging of "infinite memory" or "native multimodality." Instead, it proves — through solid architectural efficiency and real benchmark scores — that open-source models can stand at the frontier.
Sources
The following reflects information from DeepSeek's official release on 2026-04-24:
- DeepSeek V4 official launch announcement and technical notes
- Hugging Face DeepSeek organization page (model weights, MIT license)
- chat.deepseek.com / official API documentation and pricing pages
Disclaimer: Some third-party benchmark figures may change as evaluation versions are updated; refer to the latest official and leaderboard results.
Last updated: April 25, 2026