SHELFCRITTER AI Digest

MODEL

Hy3

modeltopic-noteopen-sourcechinatencent

Overview

Hy3 is Tencent‘s July 2026 flagship open-weights mixture-of-experts release — a 295B-parameter MoE with 21B active (plus a 3.8B MTP layer), 256K context, Apache 2.0-licensed, distributed as FP8 at ~300 GB on HuggingFace and free on OpenRouter through July 21, 2026. Hy3 targets the same on-desk / small-cluster inference budget as DeepSeek v4-Flash and lands as the second first-tier Chinese open-weight release in a single week alongside Zhipu AI‘s GLM 5.2 powering ZCode.

Timeline

  • 2026-07-08-AI-Digest — Tencent released Hy3 — 295B-parameter MoE with 21B active (plus 3.8B MTP layer), 256K context, Apache 2.0, distributed FP8 at ~300 GB on HuggingFace and free on OpenRouter through July 21. Simon Willison ran his pelican-on-a-bicycle SVG probe in the linked note. Narrow read: Hy3’s 21B-active MoE profile is aimed squarely at the same on-desk / small-cluster inference budget as DeepSeek v4-Flash — Apache 2.0 with free OpenRouter access is genuine practitioner availability rather than a gated preview. Structural read the digest carries: pairs with the Zhipu AI ZCode launch above as the second first-tier Chinese open-weight release in one week — the “Chinese open-weights price the cheap-token tail while Anthropic and OpenAI hold the load-bearing frontier” thesis picks up two more data points this week, and the pattern the corpus has been tracking since DeepSeek v4 shipped is not a single-lab story any more.

Key Developments

  1. 295B / 21B-Active MoE with 256K Context, Apache 2.0 (July 8, 2026): The release shape — MoE with sparse activation matching DeepSeek v4-Flash’s inference budget, Apache 2.0 open weights, free OpenRouter access through July 21 — is genuine practitioner availability. The 3.8B MTP (multi-token prediction) layer is the practitioner-relevant detail that differentiates Hy3 from other 200B+-total MoE releases.

  2. Second Chinese Open-Weights Release in a Week: Pairs with Zhipu AI‘s GLM 5.2 anchoring ZCode the same day. The pattern of Chinese labs shipping frontier-class open weights inside compressed windows continues from DeepSeek v4 through GLM 5.2 to Hy3.

See also: Tencent, DeepSeek v4, GLM 5.2, ZCode, Simon Willison, MOC - Open Source Models.