Tencent's Hy3 is a near-frontier open model you can actually run
Tencent released Hy3, a 295B open MoE model under Apache 2.0 with just 21B active params. It's small enough to self-host and free to test until July 21 — here's the operator case.
Most "open" model launches are a headline number you can't actually deploy — 400 billion parameters that need a rack of GPUs to move. Tencent's newest release is the opposite kind of interesting. On July 6, 2026, Tencent released Hy3, a 295-billion-parameter model that only activates 21 billion parameters per token — small enough to run on hardware a real business can afford, under a license that lets you own it.
What actually happened
Hy3 is a Mixture-of-Experts (MoE) model: 295B total parameters, but only ~21B active on any given request, with a 256K-token context window. Per VentureBeat, it's released under the Apache 2.0 license — the commercially permissive one, no strings — and it benchmarks competitively against models several times its active size, winning on most tasks except coding. The weights are on Hugging Face and ModelScope from day one, and Tencent is offering free API access on OpenRouter through July 21, 2026 so you can test before committing anything.
The active-parameter count is the part worth internalizing. A 21B active MoE runs far cheaper than a dense frontier model at similar quality. That's the whole game for self-hosting: you're not paying to light up 295B parameters on every token, only the 21B the router picks.
Why it matters for your business
Here's the operator move: use the free OpenRouter window to run Hy3 against your actual workload — your support replies, your product descriptions, your internal summarization — and compare it head-to-head with whatever paid API you're using now. Not on benchmarks. On your tasks.
If it holds up, you now have options you didn't have last week. You can route your high-volume, non-critical work to a model you can self-host on Apache 2.0 terms, keep the expensive frontier API for the hard 10%, and stop paying premium per-token rates for work that doesn't need them. And because the weights are yours to run, no vendor can deprecate the model, change the price, or throttle you mid-quarter. Open weights aren't a philosophy here — they're a cost-and-control decision you can measure in two weeks.
Key takeaways
- Tencent released Hy3 on July 6, 2026: a 295B MoE model with only ~21B active parameters and a 256K context window, under the permissive Apache 2.0 license
- It benchmarks competitively against much larger models (weakest on coding) and is on Hugging Face and ModelScope from day one
- Free API access on OpenRouter runs through July 21, 2026 — a no-cost window to test it on your real workload
- The low active-parameter count makes self-hosting genuinely cheap; route high-volume work here and keep frontier APIs for the hard cases
Paying frontier prices for work a smaller model could do? We benchmark open models like Hy3 against your actual workload and build the routing that sends each task to the cheapest model that gets it right. Estimate what routing could save or see how we build self-hostable AI.
Sources: Tencent, VentureBeat.
- #open-weights
- #tencent
- #self-hosting
- #ai-costs
- #portability
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Illinois just made AI vendors get audited. That's your accountability floor
Illinois SB 315, the first US frontier-AI audit law, forces the biggest AI labs into annual third-party safety audits. It doesn't regulate you — it raises the floor under your vendors.
Read itDeepSeek is building its own inference chip. Inference is a commodity now
Reuters says DeepSeek is designing a custom inference chip to cut its Nvidia bill. When the cheapest lab builds its own silicon, that tells you where your AI costs are going.
Read it