The Short Answer
Tencent open-sourced Hy4 preview on 28 August 2026: a 770B-total, 49B-active mixture-of-experts model with a 1M-token context, under a genuine Apache License 2.0 with no use restrictions. The licence is real leverage. The hardware is not free, because the BF16 weights run about 1.56 TB. With ibl.ai you own all the code and the data, so adding a model is a registry entry rather than a migration.
A permissive licence changes what you are allowed to do. Your architecture decides whether you can actually do it.
What exactly did Tencent release as Hy4 preview, and when?
Tencent released and open-sourced Hy4 preview on 28 August 2026, not this week. That distance matters, because a model that has been downloadable for over two weeks has already had its weights pulled, served and benchmarked by people outside the vendor.
The published specification is 770B total parameters, 49B active per token, and a context window exceeding 1M tokens.
The model card describes a 78-layer backbone: the first layer is a dense feed-forward network, the remaining 77 use a mixture-of-experts structure with 256 routed experts and one shared expert, and each token activates the top eight routed experts plus the shared one.
A native multi-token-prediction layer for speculative decoding sits alongside it.
In a blind evaluation Tencent ran internally, with 163 experts rating outputs across 203 engineering tasks, Hy4 preview scored an average of 2.99 out of 4.00, against GLM-5.3 at 2.92 and Kimi K3 at 2.94.
That is a vendor-run internal benchmark on a four-point scale, so treat it as a claim of parity in its own class rather than an independent result.
Tencent lists the API at $0.834 per million input tokens, $2.501 per million output, and $0.042 per million on cache hits.
Is Tencent's Hy4 preview actually open-source, or only open-weight?
On the licence, it is open-source — and the reflexive assumption about Chinese releases that this supposedly overturns needs correcting first.
Permissive licensing is in fact the majority case among Chinese open-weight releases. DeepSeek V4 and R1 ship under MIT, GLM-4.6 and GLM-5.2 under MIT, and Qwen3-Coder-480B and Qwen3.8-27B under Apache-2.0.
What has become more common at the frontier tier specifically is a bespoke licence with a commercial-revenue threshold. GLM-5.3 gates above $10 billion in revenue; Kimi K3 gates above $20 million in monthly MaaS revenue.
Note what those thresholds are indexed to. They are revenue, not monthly active users and not geography. Kimi's 100-million-MAU clause is an attribution requirement rather than a cap, and the 700-million-MAU cap people tend to remember is Meta's, on a US release.
Hy4 preview sits with the permissive group, at flagship scale, which is what makes it notable at this tier. The LICENSE file is the standard Apache License 2.0, reproduced verbatim, with a Tencent copyright notice and nothing else.
No user cap, no territorial carve-out, no revenue threshold, and no separate use policy in the repository.
The GitHub repository states the same licence.
The weights are genuinely downloadable rather than API-gated, published on Hugging Face in both a default BF16 build and an FP8 checkpoint.
One precision remains. This is an open-source licence over released weights, not a fully reproducible open-source model: the training data and the full training pipeline are not published.
For a commercial deployment decision the licence is the part that binds, and this one is permissive. For a reproducibility or provenance audit, it is not the same thing as an open corpus.
What does 770B parameters mean on a mixture-of-experts model?
It means the memory number and the compute number have come apart, and headline comparisons that ignore this are measuring different things.
On a dense model, parameter count is both how much arithmetic runs per token and how many bytes stay resident in GPU memory. On a mixture of experts they separate.
Hy4 preview's 49B active parameters set the per-token compute. Its 770B total sets the memory you must keep loaded, because any expert can be routed to on any token.
Put directly: activation sparsity reduces the arithmetic per token, and does not reduce the bytes you keep resident.
So "770B open-source" and "770B dense" are not comparable quantities, and neither is "49B active" comparable to a 49B dense model — the first is cheaper to compute than its headline implies, and far more expensive to hold.
What would it actually take to self-host a 770B model like Hy4?
More hardware than most enterprises have, and the arithmetic is public.
The default BF16 repository is roughly 1.56 TB of weights. The FP8 checkpoint runs about 814 GB of files — 813.8 billion bytes across the repository — before a single token of context is allocated.
| Node | Total VRAM | FP8 (~814 GB) | BF16 (~1.56 TB) |
|---|---|---|---|
| 8x H100 | 640 GB | No | No |
| 8x H200 | 1,128 GB | Yes | No |
| 8x B300 | 2,304 GB | Yes | Yes |
The 8xH100 node has been the standard enterprise inference configuration for two years, and it cannot load either checkpoint.
Tencent's own serving documentation agrees. The SGLang recipe lists tested hardware as H200, B200, B300 and GB300, and notes that at tensor-parallel size 8 the BF16 build needs roughly 190 GB per rank, which exceeds H200 and B200 memory outright.
The vLLM recipe was tested on 16xB200 or 8xB300.
None of this makes the release less significant. It makes the honest claim narrower: for the large majority of enterprises, Hy4 preview is a model you can route to cheaply and switch to freely, not one you will casually stand up in your own rack this quarter.
Where does pricing leverage against an AI vendor actually come from?
From the credible ability to move, which is an architectural property rather than a licensing one.
A permissive licence removes the legal barrier to switching, and that is the barrier most likely to be load-bearing.
A model you may download but may not use commercially at scale gives you far less leverage than its headline suggests, which is why the Apache-2.0 finding here is the most consequential detail in the release.
But a buyer whose applications are written against one vendor's SDK, whose prompts are tuned to one model's quirks, and whose evaluation harness only runs against production, cannot move.
That buyer watches the market price fall and pays the old price, because the switching cost exceeds the saving.
This is the same dynamic behind the open-weight price floor: open weights set what closed models can charge, and only organizations that can switch collect the difference.
The same logic applies to per-seat pricing. A bill that scales with headcount rather than consumption is the wrong shape at enterprise scale, because it ignores the falling token prices entirely — the vendor's list price per user does not move when inference gets cheaper.
The practical test is one question. When a model like Hy4 preview ships, how long does it take you to run your evaluation set against it in production conditions? If the answer is measured in days, the brief's advice to revisit your vendor assumptions is actionable.
If it is measured in quarters, it is not.
How does ibl.ai turn a new open-weight model into leverage?
By making a model a configuration entry instead of a project.
With ibl.ai you own all the code and the data.
The platform runs self-hosted inside your own perimeter with full source code access, and is model-agnostic across any LLM. You can route Hy4 preview, a frontier API model and your own fine-tune side by side, and switch between them without rewriting the application layer.
It is usage-based with no per-seat pricing, so a falling token price reaches your bill directly. It can deploy anywhere: your own cloud, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.
That combination is what converts a permissive licence into an actual negotiating position. You can add a newly released model to the registry, run your evaluation set against it, and route the workloads where it wins, without a migration and without asking a vendor's permission.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY.
Related reading: the open-weight price floor is now the market's floor — why only buyers who can credibly switch capture a falling market price, and why model switching compounds.
Sources: release date, parameter counts, context window, the 163-expert/203-task blind evaluation and API pricing from Tencent's announcement; architecture, weight formats and the 814 GB FP8 repository size from the Hugging Face model card and the FP8 checkpoint; the licence text from the repository LICENSE file and the GitHub repository; tested serving hardware and per-rank memory from the SGLang recipe and the vLLM recipe; node-capacity arithmetic from this total-vs-active parameter analysis.