V4 Pro is the clear upgrade: 1M-token context (vs 128K), stronger reasoning, and — despite being 2.4× larger — cheaper long-context inference thanks to the CSA/HCA attention stack (27% of V3.2's FLOPs at 1M tokens). V3.2 remains relevant purely on price and hosting ubiquity. New builds should start on V4; existing V3.2 deployments migrate when context length or quality ceilings bite.
DeepSeek V4 vs V3.2 at a glance
| DeepSeek V4 | DeepSeek V3.2 | |
|---|---|---|
| Vendor | DeepSeek (China) | DeepSeek (China) |
| Open weights | Yes — downloadable | Yes — downloadable |
| License | MIT — weights free to download, self-host, fine-tune, and use commercially | MIT — weights free to download, self-host, fine-tune, and use commercially |
| Parameters | 1.6T (49B active) | 671B (37B active) |
| Context window | 1M tokens | 128K tokens |
| Modalities | text | text |
| Pricing | $1.74 in / $3.48 out per 1M tokens | $0.28 in / $0.42 out per 1M tokens |
| Released | 2026-04-24 | 2025-12-01 |
Specs and pricing verified July 2026.
About DeepSeek V4
DeepSeek V4 Pro is the flagship of DeepSeek's V4 family, released on April 24, 2026, and the strongest argument yet that open-weight models compete at the frontier. It is a 1.6-trillion-parameter mixture-of-experts model that activates just 49B parameters per token, pairs a 1M-token context window with up to 384K tokens of output, and ships under a plain MIT license — no usage thresholds, no acceptable-use gate.
The headline engineering story is the attention stack. V4 combines Compressed Sparse Attention (CSA), which compresses KV entries 4× along the sequence with softmax-gated pooling, with Heavily Compressed Attention (HCA) at 128× compression. The practical result: at 1M-token context, V4 Pro needs roughly 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2. Long context stopped being a luxury and became the default across DeepSeek's services.
Full specs, benchmarks, and hardware guidance: the DeepSeek V4 page.
About DeepSeek V3.2
DeepSeek V3.2 was the model that introduced DeepSeek Sparse Attention (DSA) to production, cutting long-context inference cost dramatically while holding benchmark parity with V3.1-Terminus. Through early 2026 it was the workhorse open-weight model for coding platforms and hosting providers, scoring around 70% on SWE-bench Verified under an MIT license.
With the V4 family's release in April 2026 it moved from flagship to value tier, and it remains a well-understood, widely-hosted target with mature vLLM and SGLang support — often the pragmatic choice when V4 Pro is overkill.
Full specs, benchmarks, and hardware guidance: the DeepSeek V3.2 page.
Choosing between them
Choose DeepSeek V4 for:
- Frontier-level reasoning and coding on an open license
- Very long documents and repositories (1M-token context as the default)
- Teams that need Anthropic/OpenAI API compatibility with open-weight economics
Choose DeepSeek V3.2 for:
- Proven, widely-hosted open-weight coding capability
- Teams standardized on the V3-generation deployment stack
- Budget API workloads on the DeepSeek platform
Frequently asked questions
Is DeepSeek V4 better than DeepSeek V3.2?
V4 Pro is the clear upgrade: 1M-token context (vs 128K), stronger reasoning, and — despite being 2.4× larger — cheaper long-context inference thanks to the CSA/HCA attention stack (27% of V3.2's FLOPs at 1M tokens). V3.2 remains relevant purely on price and hosting ubiquity. New builds should start on V4; existing V3.2 deployments migrate when context length or quality ceilings bite.
Which is cheaper: DeepSeek V4 or DeepSeek V3.2?
DeepSeek V4: $1.74 in / $3.48 out per 1M tokens. DeepSeek V3.2: $0.28 in / $0.42 out per 1M tokens. Both are open-weight, so self-hosting costs depend on your hardware and utilization.
Can I self-host DeepSeek V4 and DeepSeek V3.2?
DeepSeek V4: yes — weights are published under MIT. DeepSeek V3.2: yes — weights are published under MIT.