2026-07-06

2026-07-06-tencent-hy3-open-coding-moe

agentsmodels

Summary

Hy3 (Tencent Hunyuan 3) is a 295B-parameter Mixture-of-Experts model with 21B active parameters, 256K context, Apache-2.0, fully open-weights (HF + ModelScope + GitCode). It’s a fused fast/slow-thinking reasoning + agent model. Tencent’s first major release after rebuilding its pre-training/RL infrastructure in early 2026. Vendor-reported coding/agent benchmarks: SWE-bench Verified 74.4%, Terminal-Bench 2.0 54.4%, plus search-agent scores (BrowseComp 67.1, WideSearch 70.2) — positioned as “competitive with GLM-5, closing ground on Opus 4.6 / GPT-5.4.” The Hy3-preview weights open-sourced April 23, 2026; the canonical tencent/Hy3 HF repo is dated 2026-07-02 (GA-over-preview bump or re-pointer — not independently disambiguated).

Surfaced via list-don’t-query on the HF trending index + the createdAt one-field check (07-02, fresh, unlogged). Verify-don’t-trust: benchmarks are vendor-reported; no independent SWE-bench confirmation yet.

Implications

Feeds [open coding tier] and the two-clocks frame. For ~20 days the open coding clock read as “settled at GLM-5.2 (06-16), no second base contesting the head.” Hy3 is the second credible open coding base in the field — and its story is a different axis than GLM-5.2’s:

  • Efficiency, not scale. GLM-5.2 is 753B sparse; Hy3 is 295B total but 21B active — roughly GLM-5.2-class coding capability (if the vendor SWE-bench Verified 74.4 holds) at ~1/36th the active parameters. Cheap to serve (cloud / expert-parallel vLLM), which is where the open-coding competitive pressure now lives.
  • Not a local-hardware win for reference machines. 295B total ≈ >150GB even at Q4 — exceeds all three reference profiles (36/32/12+64GB). Same verdict as GLM-5.2: cloud / heavy-quant only. The efficiency is in active params (serving cost), not total footprint (local run).
  • Tracking miss note. Hy3-preview open-sourced April 23 — before GLM-5.2 — and was never logged. The “settled at GLM-5.2” frame was hiding a contender that predated it. The daily’s list-don’t-query + createdAt check surfaced it now; the miss is that it should have surfaced in April.

← all signals