batiai/Hy3-GGUF

Hy3 (Hunyuan 3.0) GGUF — Quantized by BatiAI

IQ2_M / IQ3_XXS quantization of tencent/Hy3 (Hunyuan 3.0, 295B total / 21B active MoE). Quantized directly from official Tencent BF16 weights by BatiAI — Korean-calibrated imatrix, BatiAI-signed.

Why Hy3?

• 295B parameters, only 21B active — a frontier reasoning + agentic-coding model that runs at 21B speed. The smallest of the 2026 frontier MoEs (vs GLM-5.2 753B, DeepSeek-V4 1.6T).
• Frontier benchmarks: SWE-Bench Verified 78.0, SWE-Bench Pro 57.9, GPQA Diamond 90.4, BrowseComp 84.2 — competitive with much larger models.
• Production-grade tool-calling — first-class function-calling with dedicated parsers, agentic scaffolding stability (<4% variance) — ideal for agent pipelines.
• 256K context, 192 experts (top-8) + shared expert, 80 layers, GQA.
• Apache 2.0 — and as of the official 3.0 release, no longer geo-restricted (Korea / EU / UK now fully cleared). Free for commercial use, fine-tuning, redistribution.

Quantizations

Quant	Size	Min RAM	Target	Notes
Q4_K_M	166 GB	192 GB	256GB Mac Studio / server ⭐	Highest quality — cleanest output
IQ3_XXS	106 GB	128 GB	128GB Mac Studio	Fits a 128GB Mac with context headroom

Both quants use a diverse code + English + Korean + Chinese calibrated importance matrix (imatrix) and are built with the MTP (multi-token-prediction) head pruned (--prune-layers 80) — the speculative-decoding head gives no benefit on Apple Metal and its tensors aren't imatrix-covered, so a clean 80-layer text model is the right target here.

Verified (this build): Q4_K_M produces clean, correct Python/coding output and coherent Korean. Lower-bit quants show more zh/en token leakage on Korean, so Q4_K_M is recommended when RAM allows; IQ3_XXS is the 128GB-Mac option with slightly more leakage.

⚠️ Positioning: Hy3's strength is frontier coding / reasoning / agentic tool-calling — not Korean (Tencent model, no published Korean benchmark). For Korean chat/STT on 16GB Macs, use batiai/qwen3.6-27b. Hy3 is a frontier / high-RAM tier model (like Kimi K2.6, GLM-5.1, DeepSeek-V4) for 128GB+ Apple Silicon or a workstation/server — it does not run on 16GB/64GB Macs.

BatiAI differentiation

• Direct from official Tencent BF16 (no re-quant of community GGUF).
• Korean-calibrated imatrix — calibration set includes Korean text, tuned for Korean + English quality.
• 128GB-Mac-optimized quant selection (IQ2_M fits with context headroom).
• BatiAI metadata signature + 5-gate verification (load / basic / Korean / tool-call / MoE-routing correctness).

Usage (llama.cpp)

⚙️ Hy3 (hy_v3 architecture) requires a build with hy_v3 support. Mainline merge pending (ggml-org/llama.cpp#25395); until then use a build from that PR. Ollama support will follow the mainline merge.

⚠️ Chat template: the stock Hy3 Jinja template uses .format() calls that llama.cpp's engine rejects. This repo ships a fixed template (Hy3-chat_template.jinja) — pass it with --jinja:

# download (128GB Mac → IQ3_XXS; 256GB/server → Q4_K_M)
hf download batiai/Hy3-GGUF Hy3-IQ3_XXS.gguf Hy3-chat_template.jinja --local-dir .

# chat (Apple Silicon Metal)
./llama-cli -m Hy3-IQ3_XXS.gguf -ngl 99 -c 8192 \
  --jinja --chat-template-file Hy3-chat_template.jinja \
  -p "Write a Python function for binary search."

# raw completion (no template): add -no-cnv

License — Apache 2.0

Fully permissive: commercial use, modification, redistribution — no geographic restriction (Korea / EU / UK cleared in the official Hunyuan 3.0 release). Base model © Tencent. This repo redistributes quantized weights under the same Apache 2.0 terms.

Source & citation

• Base: tencent/Hy3 (Hunyuan 3.0)
• Quantized by: BatiAI · https://flow.bati.ai

@misc{batiai-hy3-gguf-2026,
  title  = {Hy3 (Hunyuan 3.0) GGUF — Korean-calibrated quantization},
  author = {BatiAI},
  year   = {2026},
  publisher = {Hugging Face},
  url    = {https://huggingface.co/batiai/Hy3-GGUF}
}

— BatiAI · https://flow.bati.ai