Tencent's Hy3 Is a 295B Model That Runs Like a 21B, and People Are Cramming It Onto RTX Pro 6000s

TL;DR

Tencent just dropped Hy3, an open-weight 295-billion-parameter reasoning model that behaves like a 21-billion-parameter one, because it is a Mixture-of-Experts that only wakes 21B of those params for any given token. It is Apache 2.0, has a 256K context, and ships an FP8 build (tencent/Hy3-FP8) that Tencent claims trades blows with GLM-5.2 and DeepSeek-V4-Pro while being a fraction of their size. The catch is memory: even the FP8 weights are about 300GB, so the local crowd is stacking NVIDIA RTX Pro 6000 cards, 96GB each, two and four at a time, to run it at home. Here is what is real, what is vendor-reported, and what it actually takes to host.

What Hy3 actually is

Hy3 landed on July 6, 2026 as tencent/Hy3 under a clean Apache 2.0 license, not one of the community licenses with user caps that Chinese labs sometimes attach. The spec sheet: 295B total parameters, 21B active, 192 experts with the top 8 activated per token, a separate 3.8B multi-token-prediction layer, 80 layers, and a 256K native context window.

The number that matters is that 21B active. In a Mixture-of-Experts, a router picks a small subset of "expert" sub-networks for each token instead of running the whole model. Picture 192 specialists on the payroll where every token only pulls the 8 most relevant ones into the room: you house all of them, but you only pay compute for the handful who show up. So Hy3 carries 295B worth of learned knowledge and spends roughly 21B worth of compute per token. That gap is the entire reason a model this big is even in the conversation for local hardware.

How it stacks up against GLM-5.2 and DeepSeek-V4

Tencent positions Hy3 as "comparable to high-performance open models such as GLM-5.2 and DeepSeek-V4-Pro" while being far smaller. For scale, GLM-5.2 is 753B total and DeepSeek-V4-Pro is reported around 1.6T. Tencent also claims Hy3 outperforms GPT-5.5 on FrontierScience-Olympiad, a scientific-reasoning benchmark, and that on one document task with its WorkBuddy agent it finished using 47.4% fewer tokens than GLM-5.2.

The vendor-reported headline scores:

Read it honestly and Hy3's pitch is efficiency, not a new number one. We went deep on the model it is chasing in our GLM-5.2 writeup, and the interesting claim here is not that a 21B-active model tops the leaderboard, it is that it lands near a much heavier GLM-5.2 and a much larger DeepSeek while costing a fraction of the compute per token. If that holds up on your own repos, the story is cost and footprint, not a crown. These are vendor benchmarks, so the only score that counts is the one you measure yourself.

The catch: MoE saves compute, not memory

Here is the trap people walk into. The 21B-active figure describes how much math runs per token, but you still have to load all 295B parameters into VRAM, because the router might reach for any expert at any moment. You pay for the whole mansion even though you only ever occupy two rooms at a time.

So the weight footprint is set by the full model: roughly 590GB at BF16, about 300GB for the FP8 build, and around 148GB at INT4, which the deployment guides flag as too lossy for careful reasoning. On top of that, the KV cache at full 256K context adds another 80 to 120GB per concurrent stream. Tencent's own recommendation is about 8 GPUs of the H20-3e class, and cloud deploy guides land on 8x H200 (roughly 1,128GB) for BF16 with real context, or 4x H200 for FP8 at shorter context, running vLLM or SGLang with expert parallelism and pushing 2,000 to 4,500 tokens per second of aggregate throughput once MTP speculative decoding is on. That is a rental-cluster answer, not a desk answer.

Why the RTX Pro 6000 crowd cares

This is where the RTX Pro 6000 Blackwell comes in. It is the most VRAM you can buy in a single retail workstation card: 96GB of GDDR7 at about 1.79 TB/s, roughly 3x the bandwidth of an RTX 4090. NVIDIA even sells a 300W Max-Q blower variant built explicitly for people who want to bolt several into one chassis, which tells you exactly who the customer is.

One card is not enough, since the INT4 weights alone are 148GB, so the shopping list is simple arithmetic: two cards (192GB) hold a 4-bit build with room for a modest context, and four cards (384GB) hold the FP8 weights with headroom for real KV cache. And because only 21B params are active per token, once the weights are resident the decode speed is bounded by shuffling 21B worth of data, not 295B, so a multi-card box stays genuinely usable for a single user instead of crawling. For reference, one of these cards already runs a dense 122B model at 4-bit in the tens of tokens per second; spreading Hy3's active slice across a few of them keeps interactive use realistic.

The kicker is what you get for the money: a private, Apache-licensed, near-frontier reasoning model that never phones home. For anyone drawn to GLM and DeepSeek but nervous about the hosted APIs, self-hosting Hy3 sidesteps the China-data-residency question entirely, because the weights run on hardware you own. Four cards at roughly $9,000 each is a lot of money right up until you price the alternative, which is explaining to your compliance team why your codebase took a field trip through someone else's datacenter.

Should you actually build this

Be honest about who this is for. If you just want the best answer per dollar and your data can leave the building, rent a cluster or hit an API and skip all of it. If you are running many agents and you care about throughput, 8x H200 in the cloud is the boring, correct answer, and the per-token cost lands under two dollars per million.

The RTX Pro 6000 path is for a specific person: someone who wants a genuinely capable reasoning model fully under their own roof, on hardware they own, with open weights they can inspect, and who will spend workstation-GPU money to get it. At roughly $8,500 to well over $13,000 per card (NVIDIA's marketplace has listed it as high as $13,250), a four-card FP8 rig is a $40,000-plus machine: cheap next to an 8x H200 node, expensive next to an API bill. Hy3 is arguably the best current fit for that build precisely because 21B active means a home rig can keep pace. Keep the asterisk taped to your monitor anyway: the benchmarks are Tencent's, efficiency is the real headline, and your own workload is the only benchmark that pays your bills.

Key Takeaways

• What it is: Hy3 is Tencent's open (Apache 2.0) 295B Mixture-of-Experts with 21B active, 192 experts top-8, a 256K context, released July 6, 2026, with an FP8 build shipped as tencent/Hy3-FP8.
• The pitch is efficiency: Tencent claims parity with GLM-5.2 (753B) and DeepSeek-V4-Pro (reported ~1.6T) and a FrontierScience-Olympiad win over GPT-5.5, at a fraction of the size and only 21B active per token. All vendor-reported.
• MoE saves compute, not memory: you still load all 295B params, so weights run ~590GB at BF16, ~300GB at FP8, or ~148GB at INT4, plus 80-120GB of KV cache at 256K context.
• Datacenter path: Tencent recommends ~8 GPUs (H20-3e class); guides use 8x H200 for BF16 or 4x H200 for FP8, at 2,000-4,500 tok/s aggregate with MTP speculative decoding.
• Local path: stack 96GB RTX Pro 6000 cards, two for a 4-bit build and four for FP8, usable at home because only 21B params activate per token.
• The real reason to self-host: an Apache-licensed, near-frontier model that never leaves your building, with no China-hosted-API data-residency question.

Sources: Tencent Hy3-FP8 (Hugging Face), MarkTechPost, GIGAZINE, Spheron deploy guide, Hardware Corner (RTX Pro 6000), NVIDIA H200