Tencent Hy3: reading a 295B open-weight MoE without the hype

Introduction — what Hy3 actually is
What is actually worth noticing in the architecture
The open-weight landscape and the license story
How far to trust the capability claims
Closing
References

Introduction — what Hy3 actually is

On July 6, 2026, Tencent's Hy team (formerly Hunyuan) released Hy3 as open weights — the production version of the preview that shipped on April 23. Clear up one confusion first: the name looks like a 3D generator, but Hy3 is a text-based reasoning and agent large language model. It is a different line from Hunyuan3D, which produces 3D assets.

The short version is this. It is a 295B-parameter Mixture-of-Experts (MoE), but only about 21B parameters actually fire for any given token. Context length is 256K, and weights ship in BF16. The license is Apache 2.0 — no geographic carve-out, no field-of-use clause. The weights are on Hugging Face (tencent/Hy3) and ModelScope, with progressive rollout to third parties such as OpenRouter, Cline, and OpenCode.

What is actually worth noticing in the architecture

A few numbers describe this model well. There are 192 experts, of which top-8 activate per token. That produces the "295B on disk, 21B in compute" shape. The inference compute is close to a 21B dense model, but you still have to hold all 295B in memory. Even by plain arithmetic the weights alone are roughly 590GB (2 bytes x 295B) in BF16, which is why Tencent recommends "8 GPUs, H20-3e or larger memory." It is worth stating plainly that the recommended card is the China-market H20 family.

The key specs, gathered in one place:

295B total parameters / 21B active (192 experts, top-8)
3.8B Multi-Token Prediction (MTP) layer for speculative decoding
256K context, BF16 precision
Attention: 64 heads (GQA, 8 KV heads, head dim 128)
Reasoning control reasoning_effort: no_think (default), low, high
Serving stack: vLLM and SGLang; quantization via the AngelSlim toolkit

The 3.8B MTP layer feeds speculative decoding in vLLM and SGLang to speed up token generation. The reasoning switch is equally pragmatic: easy requests answer directly under the default no_think, and only hard problems get flipped to high for a long chain of thought — so you do not pay latency and tokens for reasoning you did not need.

A word on practical operation. Because you have to load roughly 590GB, this is a model you serve on a server, not run on a laptop. Low-bit quantization via AngelSlim cuts the memory requirement enough that trying it on a multi-GPU workstation becomes plausible — and the 21B active-parameter count works in your favor again here.

The open-weight landscape and the license story

Hy3 did not appear in a vacuum. It is the latest entry in the Chinese open-weight wave that runs through DeepSeek, Zhipu's GLM, and Alibaba's Qwen. What this camp shares is releasing near-flagship capability under permissive licenses.

Here the choice of Apache 2.0 matters more than it looks. Chinese "open" models have often shipped under bespoke community licenses carrying usage or regional clauses — the Hy3 preview itself did. The production Apache 2.0 carries none of that, which sharply lowers the adoption barrier for any organization that has to run a legal review first.

That is why, for a Korean developer, the license history is the genuinely interesting part. According to multiple outlets, the April preview shipped under a restrictive "Tencent Hy Community License" that explicitly excluded the EU, the UK, and South Korea from its usage terms. The July release drops that clause entirely and switches to Apache 2.0. This history comes from secondary coverage rather than the primary announcement or model card, so I flag the provenance. But if accurate, South Korea went from "not allowed to use it" to "unrestricted commercial use" within a few months.

How far to trust the capability claims

The published benchmarks (all vendor-measured) look impressive:

GPQA Diamond 90.4
SWE-Bench Verified 78.0 / SWE-Bench Pro 57.9
HLE (Humanity's Last Exam) 53.2
USAMO 2026 72.0 / IMOAnswerBench 90.0
Deep SWE 28 / Apex Agents 25.6

Tencent frames it as "intelligence comparable to flagship models two to five times its parameter scale." It claims to be "on par with" GLM-5.2 and DeepSeek-V4-Pro while being far smaller than those 753B and 1.6T rivals, and reports processing documents with 47.4% fewer tokens than GLM-5.2.

It is also worth noting that several of these — Deep SWE, Apex Agents — are relatively new agentic suites without many published baselines from rival models, so a high-looking score is hard to place in context.

This is where a cool head helps. Every one of these numbers is vendor-reported, and as of writing no independent third party has verified them. Even inside the vendor's own data the weak spot shows: on coding it trails GLM-5.2 — SWE-Bench Verified 78.0 versus 84.2. The claim of beating GPT-5.5 is a categorical statement about one task, FrontierScience-Olympiad, with no score table attached. On the other hand, the fact that low results like Deep SWE 28 and Apex Agents 25.6 appear on the model card at all is mildly reassuring.

Closing

The most certain thing about Hy3 is not its benchmark ranking but its accessibility. With 21B active parameters inference is cheap, Apache 2.0 removes legal friction, and it runs on the standard stack (vLLM, SGLang) out of the box.

The benchmarks deserve to be read as vendor claims until verified. But the plain fact that there is now one more "good enough and unrestricted" open model needs no verification. Actual adoption will be settled not by a leaderboard but by a few days of running it on your own workload.