LIFEHUBBER
Theme

AI Resources

Hy3

Hy3 is a Tencent Hy Team Mixture-of-Experts language model for long-context reasoning, coding, document work, tool use, and agent-style workflows.

The official materials describe Hy3 as the follow-up to Hy3 Preview, with 295B total parameters, 21B active parameters, 256K context length, public Hy3 and Hy3-FP8 model weights, Apache-2.0 licensing, and vLLM or SGLang deployment notes. Use this as a first read, not a recommendation. Open the original project before trusting details like terms, limits, privacy, cost, setup, or safety.

What it is

Large MoE language model

Tencent presents Hy3 as a 295B-parameter MoE text-generation model with 21B active parameters, a 256K context window, BF16 weights, and an FP8 variant for readers comparing large public model releases.

Why it stands out

Agent, coding, and reliability focus

Tencent says the final Hy3 release builds on Hy3 Preview with feedback from 50+ product teams and changes aimed at tool calling, output formats, hallucination reduction, long-context behavior, coding, and agent workflows.

Availability

Public weights with technical serving paths

The Hugging Face and GitHub materials link the main Hy3 weights, an FP8 quantized model, ModelScope, GitCode, CNB, a finetuning guide, and serving recipes for vLLM and SGLang.

Why it matters

What makes it useful

Hy3 helps readers compare a large public model at the model layer before choosing a chat app, agent scaffold, or serving stack. Tencent connects the release to 256K context, reasoning modes, tool-call parsing, vLLM and SGLang recipes, and project-reported coding and agent evaluations, so the practical question becomes which setup can actually run and review the model well.

Notable points

What stands out

Tencent describes Hy3 as a final release after Hy3 Preview, with feedback from 50+ product teams, a blind evaluation involving 270 experts and 312 valid comparisons, and internal evaluations around hallucination, commonsense errors, and multi-turn intent tracking. Treat those performance details as Tencent-reported claims until tested in the reader's own workflow.

Before using

What to review

The Hugging Face model card, GitHub README, license file, and serving recipes before relying on access, redistribution, benchmark, or deployment details.

Hardware and serving expectations, since Tencent says serving the full model on eight GPUs requires H20-3e or other large-memory GPUs.

Whether BF16, FP8, vLLM, SGLang, a hosted API, or a different model family fits the actual task, budget, latency, and data-handling needs.

The Hugging Face page status for hosted inference, because the model page currently shows no listed Inference Provider deployment.

Reader fit

Who may find it relevant

Readers tracking large public language models for coding, reasoning, tool use, and long-context work.

Builders comparing model weights that can sit underneath agent scaffolds, serving frameworks, and OpenAI-compatible local API routes.

Teams that want to separate model choice from the app or agent shell around it before committing a workflow.

Less relevant for readers looking for a small local model, a no-setup chatbot, or Tencent Hunyuan 3D/image-to-3D generation.

Editorial note

Why LifeHubber lists it

LifeHubber lists Hy3 because it turns the earlier preview page into a current checkpoint for comparing large agent-oriented model releases. A reader can check the final weights, FP8 route, serving recipes, and Tencent-reported workflow evaluations before building around one model stack.

Source links

Source materials

Reader note

Before relying on this entry

LifeHubber lists entries to help readers inspect AI projects, not to endorse them or prove they are safe, suitable, accurate, maintained, or right for a specific use. We do not verify every entry in depth. Before relying on anything listed, review the original materials, terms, privacy practices, limits, and risks that matter for your situation.

Sponsored

Sponsored

Related in LifeHubber

Keep the thread going

Follow the next layer with AI Resources for AI projects with original links and practical caveats, AI Guides for decision habits for messy AI choices, AI Access for free and low-cost ways to compare AI model access, AI Ballot for a clearer view of what readers are leaning toward, and AI Radar for AI stories that deserve a second look.