Hy3
Tencent·
LLMsopen-weightMulti-GPU Self-Host
Overview
Tencent's Hunyuan 3 flagship: a 295B-parameter MoE activating 21B per token (192 routed experts with top-8 routing plus an always-active shared expert), 256K context window and three selectable reasoning-effort modes. Full release under Apache 2.0 — unlike the April preview license, without territorial restrictions.
Capabilities and innovations
295B total / 21B active MoE256K token context windowThree reasoning-effort modesApache 2.0 licenseDense-MoE hybrid with always-active shared expertMulti-Token-Prediction layers (3.8B)
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 90.4% | vendor GPQA Diamond 90.4, vendor model card. |
| SWE-bench Verified | 78% | vendor SWE-bench Verified 78.0, vendor model card; scaffolding variance within 4% across CodeBuddy, Cline and KiloCode. Era-1 benchmark, recorded for continuity. |
| SWE-bench Pro | 57.9% | vendor SWE-bench Pro 57.9, vendor model card. |
- Humanity’s Last Exam (no tools): independent evaluation pending — Vendor reports HLE 53.2 with tools only — not comparable to the no-tools values tracked here; independent no-tools run pending.
- MMLU-Pro: independent evaluation pending — Not in the vendor card; independent per-benchmark backfill planned.
- Terminal-Bench 2.x: not reported by the vendor
- MMMU-Pro: not reported by the vendor — Text-only model.
Architecture and hardware
- Parameters
- 295B total, 21B active per token (MoE)
- Estimated VRAM at Q4
- ~170 GB, Multi-GPU Self-Host class
- Recommended runtime
- vLLM
- License
- Apache 2.0
Links
More from Tencent
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.