- 🆕 Tencent Hy Team's 295B Mixture-of-Experts model, 21B active per token.
- 📏 256K-token context window for long, complex inputs.
- 🧠 Built for reasoning, STEM and instruction following.
- 🔧 Adds a Multi-Token-Prediction layer for speculative decoding.
- 🎯 Targets coding, agents, function calling and context learning.
- ⚡ MoE routing keeps inference cost near a 21B-active model.
- 🌐 Open weights on Hugging Face under Tencent's Hy community license.
- 📚 Separate base checkpoint also published.
Tencent is a Chinese multinational technology conglomerate headquartered in Shenzhen, and one of the world's highest-grossing multimedia companies by revenue. Best known for its sprawling presence in social platforms, digital services, and as the largest player in the global…
Explore 2 more models by Tencent →Hy3 Preview is an open-weight large language model from Tencent's Hy team, released in 2026. Architecturally it is a Mixture-of-Experts model with 295B total parameters, activating only 21B per token, paired with a Multi-Token-Prediction layer that supports speculative decoding. Per the model card and an accompanying Hugging Face technical write-up, it uses a routed-expert design with a shared expert, grouped-query attention, and a 256K context window, distributed in int4 quantization.
According to Venice's catalog description, the model is aimed at complex reasoning, instruction following, in-context learning, coding and agent tasks. The Hugging Face write-up frames the architecture as a "rebuilt" Hunyuan with a new reasoning recipe, routing routine queries quickly while directing harder problems to deeper reasoning chains. Its listed capabilities include reasoning, code optimization, function calling and web search.
As the newest entry in Tencent's Hy family, the team presents Hy3 Preview as an upgrade over earlier Hunyuan releases, with the MoE design intended to deliver larger-model quality at a smaller active-parameter inference cost. Tencent also reports results on STEM and agent evaluations, but these are presented as self-reported figures rather than independently verified numbers.
The weights are published on Hugging Face, with a separate base checkpoint available for further fine-tuning, under Tencent's Hy community license. The release has drawn meaningful download and engagement activity on Hugging Face since launch.
This About section is AI-generated from public sources via VeniceStats + Venice inference, with no human editing. It may contain inaccuracies.
Daily usage · Hy3 Preview · Organic sellers
| Seller | Rep↓ | Sales | Tokens | Routing | Input $/M | Cached $/M | Output $/M | Save |
|---|---|---|---|---|---|---|---|---|
| Apex TEE Test 0xe672…7955 | 0.00 | — | — | gated | $0.0756 | $0.0249 | $0.252 | −58% |
"Best price" and the seller table are live Antseed catalog data (advertised $/1M tokens — or $ per generated image for unit-billed image models — not settled amounts). Reputation = buyer trust score (0-100, the Antseed SDK's own formula). "Routing" = the SDK's default buyer routing (what the VPR desktop app ships with): a trust ≥ 60 gate on the effective reputation, then cheapest-first among routable sellers; live failover state (per-peer cooldowns) is buyer-side runtime and not included. Model knowledge (TLDR, provider, About) via the VeniceStats enrichment layer. Advertised catalog, not the model used in any specific purchase. "Usage on Antseed" counts only settlements whose buyers share the per-model split on-chain (metadata v2/v3, opt-in), so every usage figure is a lower bound.