DeepSeek V4 vs Hy3: which large open model should you use?
DeepSeek V4 is MIT-licensed with a 1M-token context in 1.6T and 284B sizes; Hy3 is one 295B Apache 2.0 model with 256K. Compare size, context and license.
Verdict. Both are large open mixture-of-experts models for agent and coding work, but they differ in size, context and license terms. DeepSeek V4 is MIT-licensed, has a 1M-token context and comes as a 1.6T Pro with 49B active parameters and a 284B Flash with 13B active, and V4-Flash-0731 is about 167 GB. Hy3 is a single 295B model with 21B active parameters, a 256K context, Apache 2.0 terms and no territory limits, and it needs about 598 GB in BF16 or 300 GB in FP8. Both are text-only, and the benchmark tables of both developers are their own.
Specifications
| Specifications | DeepSeek V4 | Hy3 |
|---|---|---|
| Developer | DeepSeek | Tencent |
| Released | April 24, 2026 | July 6, 2026 |
| Total parameters | 1.6T (V4-Pro-0813) | 295B (Hy3 (BF16)) |
| Active parameters | 49B | 21B |
| Architecture | Mixture of experts | Mixture of experts |
| Context window | 1M tokens | 256K tokens |
| Input | Text | Text |
| License | MIT (Permissive) | Apache 2.0 (Permissive) |
| Memory for the smallest variant (4-bit) | 176 GB (V4-Flash (preview)) | 204 GB (Hy3 (BF16)) |
Performance claims come from the developer and have not been independently verified.
Pick DeepSeek V4 if
- You need up to 1M tokens of context in one session.
- You want a bigger step up with the 1.6T Pro, or a cheaper step down with the 284B Flash and its 13B active parameters.
- You want the plain MIT license.
- You have a multi-GPU server or a very large unified-memory machine, since even the Flash is about 167 GB.
Pick Hy3 if
- 256K tokens is enough and you want a single 295B model with 21B active parameters.
- You want Apache 2.0 terms with no territory limits.
- You plan to fine-tune or run reinforcement-learning experiments, since Tencent publishes the post-training instructions.
- You can make sure you download the final release and not the April preview, which used a territory-limited license.