Hy3 — Tencent drops the restrictions and ships a real Apache 2.0 flagship

Tencent released Hy3 on July 2 — a 295B-total / 21B-active Mixture-of-Experts flagship under a clean Apache 2.0 license, with 256K context, BF16 and FP8 checkpoints, and day-one vLLM/SGLang support. The license is the headline: April's Hy3-Preview geo-blocked the EU/UK/South Korea and carried a 100M-MAU commercial cap; the full release drops all of it. At ~300 GB even in FP8 it is not a consumer-GPU model — but as open flagships go, this is one of the cleanest licensing stories of the year.

Verdict: Tencent's flagship goes fully Apache 2.0 — no geo-blocks, no MAU caps — but 295B total keeps it big-iron, not a single-box local pick

The take

The facts, verified against the Hugging Face card via the HF API (`tencent/Hy3`, created 2026-07-02, license tag apache-2.0, safetensors total 298.8B including a 3.8B multi-token-prediction layer): 295B MoE, 192 experts with top-8 routing, 21B active per token, GQA, 256K context, an FP8 variant two days later. Traction is real — ~740 likes and 9K+ downloads in eleven days. Vendor benchmarks put it at parity with open flagships "2–5× its size" (SWE-bench Pro 57.9, Terminal-Bench 2.1 71.7) and claim a halved hallucination rate; treat all of that as vendor numbers until independently reproduced.

Why it matters: the permissively-licensed frontier keeps getting more crowded — Hy3 joins GLM-5.2 (MIT), LongCat-2.0 (MIT), Kimi K2.7 (Modified-MIT), and DeepSeek V4 (MIT) in the big-iron open tier, and it is the first of that group to ship restriction-free Apache 2.0 from day one after a restricted preview. For anyone building on open weights commercially, the license delta between "Apache 2.0, no asterisks" and community licenses with MAU caps is the difference between usable and not.

Our call: no planner-pick change. 295B total is datacenter or serious-multi-GPU class — FP8 needs ~300 GB, and even aggressive community quants will not fit a 128 GB box with usable context. The 21B-active design makes hosted inference cheap and fast, which is where most people will touch it (Tencent cloud, OpenRouter as availability spreads). We track it as a frontier comparator; local coding and chat picks are unchanged.

Where this fits

Models: GLM-5.1 · Kimi K2.6 · DeepSeek V4-Pro · Qwen 3.6-35B-A3B

Hardware: NVIDIA DGX Spark · Dual RTX 5090 · Mac Studio M3 Ultra 96 GB

Sources

Next step

Try this in the planner→