clauding.de
DE
• 2 min read AI-generated

Tencent Hy3: 295 Billion Parameters, Apache License, and Cheaper Than Anything From OpenAI

Copy article as Markdown

Tencent just released Hy3, an open-source MoE model that competes with models two to five times its size on many benchmarks. And it costs almost nothing to run.

Tencent Open Source MoE LLM China
Featured image for "Tencent Hy3: 295 Billion Parameters, Apache License, and Cheaper Than Anything From OpenAI"

Tencent released Hy3 last week — a Mixture-of-Experts model with 295 billion total parameters, of which only 21 billion are active per token. The result: a model that competes with GPT-5.5 and Claude Opus on paper, but needs a fraction of the compute.

What Hy3 can do

The architecture uses 192 experts with top-8 routing. That means only 8 of the 192 experts fire for each token — the rest stay dormant. This slashes inference costs dramatically, which is why Tencent can price the API at 1 yuan per million input tokens.

The context window is 256K tokens. On benchmarks for science, math, and general reasoning, Hy3 beats GPT-5.5 according to Tencent — though it falls behind Z.ai’s GLM-5.2 on coding, which has more than twice the parameters.

What stands out: since the April preview, Tencent brought hallucination rates down from 12.5% to 5.4%, and commonsense errors dropped from 25.4% to 12.7%. That sounds like serious internal testing — Tencent says 50 product teams used the model before launch.

Why this matters

Hy3 ships under Apache 2.0. That means commercially usable, self-hostable, fine-tunable — without the geographic restrictions that some earlier Chinese open-source models carried.

OpenRouter already lists a free route (tencent/hy3:free) running until July 21.

For the market, this means: after DeepSeek and GLM-5.2, Tencent is now delivering a near-frontier model under an open license. The cost advantages of Chinese open-source models are becoming a real problem for Anthropic and OpenAI — as CNBC’s investigation from yesterday showed, 30 to 46 percent of US enterprise API tokens are already flowing to Chinese models.

My take

Hy3 isn’t a DeepSeek moment — it lacks the surprise factor. But it shows the wave of Chinese open-source models wasn’t a one-time event. Tencent, Z.ai, Meituan — they’re all shipping models that are ‘good enough’ for most enterprise tasks while costing 60 to 90 percent less than US competitors. The question is no longer whether companies will use these models, but how policymakers will respond.

Sources: