Tencent released Hy4 preview on August 28 with 770 billion total parameters, 49 billion active per token, and a context window beyond one million tokens. The checkpoint is available in BF16 and FP8 under Apache 2.0, accompanied by inference code and a fine-tuning pipeline.
That combination makes Hy4 legally simpler than several open-weight flagships released this month, though the model is difficult to run. Tencent’s own serving recipe uses eight-way tensor parallelism for FP8, and the API costs $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached tokens. Most buyers should start with the API and make self-hosting prove its case.
A large MoE designed around long agent runs
Hy4 uses 78 transformer layers, 256 routed experts, and one shared expert. It activates eight routed experts plus the shared expert for each token. A native multi-token-prediction module adds another ten billion total parameters, 0.7 billion active, to predict several future tokens and improve decoding efficiency.
The attention architecture combines Tencent’s Gated DeepSeek Sparse Attention with IndexCache. Sparse attention reduces the amount of historical context each token must examine, while IndexCache is intended to reuse the indexing work across layers. An improved hyper-connection design, iHC, gives the model additional residual paths. All three choices address the same problem: a million-token context is only useful when the serving stack can process it without turning every request into an infrastructure event.
Tencent positions Hy4 for coding, office work, and scientific research. The public checkpoint is text-only, despite a product story that spans documents and presentations. Any spreadsheet or slide workflow therefore depends on the agent’s tools and document parsers rather than native visual understanding.
The FP8 serving example uses vLLM or SGLang with a tensor-parallel size of eight. That is a useful reality check. A 49-billion-active model may have a lower compute cost per token than a dense 770B model, but the entire checkpoint still has to live somewhere. Even after quantization, this is data-center deployment rather than a large workstation model.
The headline comparison is an internal preference test
Tencent reports a blind evaluation with 163 of its own experts across 203 engineering tasks. Hy4 averaged 2.99 out of 4, compared with 2.92 for GLM 5.3 and 2.94 for Kimi K3. In Tencent’s detailed breakdown, Hy4 won 46.8% of comparisons against GLM 5.3, tied 12.8%, and lost 40.4%. Against Kimi K3 it won 51.2%, tied 7.9%, and lost 40.9%.
This is more informative than a claim of universal superiority, but it remains an internal test. Tencent selected the tasks, assembled the evaluators, and released the result. A difference of 0.05 or 0.07 on a four-point rating scale should be treated as evidence that the models are competitive, not that Hy4 has established a decisive lead.
The preview label carries weight here. Tencent’s model card acknowledges that Hy4 can spend too long reasoning and may over-verify completed work. Those behaviors matter in an agent because they inflate token use and can make a fast model slow at the task level. The right evaluation should record total task cost, elapsed time, tool errors, and unnecessary actions alongside pass rate.
Tencent also says Hy4 participated in its own development by proposing and testing changes to training methods, data, evaluations, and low-level operators. The company attributes a 31.8% improvement in end-to-end inference throughput to model-assisted work on operator fusion and communication. No baseline configuration or reproducible experiment accompanies the newsroom claim, so it is better read as a case study from Tencent’s development process than as a model benchmark.
Apache 2.0 is a competitive feature again
The repository’s Apache 2.0 license is uneventful, which is precisely the advantage. There is no revenue trigger, model-as-a-service restriction, separate-license clause, or product-category gate. Companies can host the weights, modify them, offer them through an API, and build commercial coding tools under familiar terms.
That is a material contrast with Qwen3.8-Flash-Next, whose public license requires separate permission for commercial MaaS and standalone coding or office assistants. It also differs from GLM 5.3’s license split, where the full model places a security-review condition on MaaS operators above a very high revenue threshold while Flash remains MIT.
For model vendors, Apache licensing is not charity. It lowers the cost for inference providers to add Hy4, increases the number of downstream optimizations, and gives enterprise buyers an alternative when another lab’s custom terms are hard to approve. Tencent still earns API revenue from organizations that do not want to operate a distributed 770B checkpoint.
The practical lock-in question therefore moves from law to systems engineering. A team can leave Tencent’s API without negotiating permission, but only if another provider supports Hy4 or the team can run the serving stack itself. The Apache license removes a legal barrier, not the migration cost.
API first, weights when control justifies the rack
The launch pricing is aggressive enough that an API pilot should be the default. At $2.501 per million output tokens, a team can run a representative coding or document benchmark before reserving hardware. Cache hits at $0.042 per million also make long, repeated contexts cheaper on the managed service, the exact usage pattern Hy4’s million-token window targets.
Self-hosting becomes reasonable when data location, fixed high utilization, custom fine-tuning, or service continuity justifies the operational burden. The calculation must include KV-cache memory at the intended context length, not just the weight files. It should also include distributed-inference failures, observability, model upgrades, and the engineering time needed to reproduce Tencent’s serving configuration.
Hy4 combines a current large-agent architecture with plain Apache terms. Its performance case is less settled. The most prominent comparison comes from Tencent, the model is explicitly a preview, and the documented tendency to over-reason can erase some of its token-level efficiency.
The weights provide a credible exit from the API. They are not a reason to begin with a rack. Test the managed model against the actual workload, retain the Apache checkpoint as leverage, and move only when control is worth more than the infrastructure Tencent is subsidizing.
Sources
- Tencent Releases and Open-Sources Tencent Hy4 preview — Tencent
- Tencent Hy4-preview model card — Tencent on Hugging Face
- Tencent Hy4-preview repository — GitHub
- Hy4-preview License — GitHub