Model
Model explorer

Hy3

OPEN
Tencent · Hunyuan family · released Jul 6, 2026

Full release of Tencent's Hunyuan 3 (upgraded from an April 2026 preview build), Tencent's first genuinely unrestricted Apache 2.0 flagship. Newer than Hunyuan-A13B already in corpus (dated 2025-06-27). Pricing converted from RMB 1/RMB 4 per million input/output tokens at ~7.1 RMB/USD.

ReasoningCodingVisionFunction callingTool useAgentic
2380.6
Elo · rank #52
Parameters
295B
Active params
21B (MoE)
Context
256K tokens
Architecture
Mixture-of-Experts hybrid fast/slow-thinking model, 295B total / 21B active
License
Apache 2.0
Languages
—
API price (in/out)
$0.14 / $0.55
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
Arena EloHuman preference1454#37
best: Claude Fable 5 · 1504
BrowseCompAgents84.2%#12
best: Kimi K3 · 91.2%
GPQA DiamondReasoning90.4%#31
best: GPT-6 Astra · 96.0%
Humanity's Last ExamReasoning37.0%#38
best: Claude Opus 5.5 · 67.7%
SWE-bench ProCoding57.9%#26
best: Claude Opus 5.5 · 89.9%
SWE-bench VerifiedCoding78.0%#38
best: Claude Opus 5 · 97.0%
Terminal-Bench 2.0Coding71.7%#28
best: DeepSeek-V4.1-Flash · 90.6%
USAMOMath72.0%#3
best: Claude Opus 4.8 · 96.7%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
—
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
—
Fine-tune it
Permissive
QLoRA200.6 GB2× H200 141GB
LoRA628.4 GB4× B200 192GB
Full fine-tune4734.8 GBbeyond 8× B200
QLoRA SFT on ~10k samples ≈ $10.86 (2× H200 141GB)
Plan a fine-tune →
Hunyuan family
Elo progression across releases
Hunyuan (1st-gen)Sep 20231000.0Hunyuan-Large (A52B)Nov 2024980.9Hunyuan-TurboSFeb 20251226.6Hunyuan-T1Mar 20251571.3Hunyuan-A13BJun 20251474.3Hy3Jul 20262380.6
Resources
HFtencent/Hy3DOCShttps://www.tencent.com/en-us/articles/2202386.htmlAPIAvailable via Tencent API
API price $0.14/$0.55 · each benchmark row carries its own source badge (see methodology)