Tencent

Hy3

Name: Hy3
Rating: 84 (8 reviews)
Author: Tencent

Frontier

HuggingFace

18.6KDownloads871LikesJul 2026Veröffentlicht262K TokenKontextApache 2.0Lizenz91 HerausragendQualität

Hy3 (295B parameters) requires approximately 186.3 GB of VRAM with Q4_K_M quantization. As a Mixture of Experts model with 21B active parameters, it uses less memory than its total parameter count suggests. For the best balance of quality and speed, we recommend hardware with at least 215 GB of VRAM.

Loslegen

— kopieren & einfügen, um lokal auszuführen

Copy-paste commands to run Hy3 on your machine.

Run

docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \
  --hf-repo "tencent/Hy3" \
  --hf-file "Hy3-Q4_K_M.gguf" \
  -c 4096 -ngl 99

Quick specs

Parameters295B (21B active)

Architecturemoe (MoE)

Context262K tokens

Modalitytext

Min RAM115.1 GB

Rec. RAM180 GB (Q4_K_M)

LicenseApache 2.0

FamilyHunyuan

✓ Chat✓ Reasoning

About this model

Hy3 is Tencent's 295B-parameter Mixture-of-Experts flagship with 21B active parameters per token (plus a 3.8B MTP layer), built on feedback from 50+ Tencent products after the Hy3 Preview release. It uses 192 experts with top-8 routing across 80 layers and a 256K context, and is positioned to rival open-source flagships with 2-5x its parameter count.

•295B total / 21B active per token — 192 experts, top-8 routed, plus 1 shared expert.
•Rivals flagship open-source models with 2-5x more parameters.
•256K context with strong multi-turn intent tracking (internal issue rate cut from 17.4% to 7.9%).
•Apache 2.0 licensed.
•Tencent recommends 8x H20-3e or similar high-memory GPUs for full-precision serving.

Schnellauswahl

Bestes BudgetS

AMD Instinct MI350X 288GB~$8,000 — 99 tok/s

Beste Hardware

Top-Empfehlungen für Hy3

AMD Instinct MI350X 288GBS

288 GB

AMD Instinct MI325X 256GBS

Dieses Modell ausführen

Hy3 on AMD Instinct MI350X 288GB Hy3 on AMD Instinct MI325X 256GB Hy3 on NVIDIA GB200 192GB

Quantisierungsoptionen

VRAM-Schätzungen nach Quantisierungsstufe

No hardware detected — fit column shows raw VRAM estimates

Quant	Bits	VRAM	Quality	Fit
Q1_0_G128	1.125	42.5 GB	Very Low	—
Q2_0_G128	1.71	78.8 GB	Low	—
Q2_K	2	115.1 GB	Low	—
Q3_K_S	3	144.6 GB	Low	—
NVFP4	4	165.2 GB	Medium	—
Q4_K_M	4	180.0 GB	Medium	—
Q5_K_M	5	212.4 GB	High	—
Q6_K	6	241.9 GB	High	—
Q8_0	8	315.7 GB	Very High	—
F16	16	604.8 GB	Maximum	—

Quality benchmarks

Hy3 benchmark scores

Benchmark verified

Coding

SWE-bench Verified78.0%

HumanEval+—

Aider Polyglot—

LiveCodeBench—

Reasoning

MMLU-Pro—

GPQA Diamond90.4%

MATH-500—

ARC Challenge—

Source: official · 2026-07-02

Hardware-Kompatibilität

Eignungsschätzungen für alle Hardware

Rechner öffnen

Computing compatibility...

Speicheraufschlüsselung

Reference: RTX 2060 6GB

Weights180.0 GB

KV Cache4.9 GB

Runtime0.9 GB

Headroom0.6 GB

Häufig gestellte Fragen

FAQ — Hy3

Siehe auch

Quantisierungsleitfaden Bewertungsmethodik Rechner öffnen