Hy3-preview NEW
Tencent Hunyuan's open MoE reasoning model
Tencent's open-weight reasoning model in the Hunyuan 3 line. 295B total parameters with 21B active per token (sparse mixture of experts), 256K context, and reasoning-style answers. Roughly the open-weights peer to DeepSeek V3.2 by AA Intelligence Index.
Intelligence
Good
Speed
Slow
81 tok/s output
Cost
Low
$0.07 in / $0.26 out
Context
256K
Up to 256,000 tokens
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on
Artificial Analysis. Cost buckets are fixed dollar
thresholds keyed off output-token price ($/M out).
Intelligence
- Top 1%≤ 1%
- Top 5%≤ 5%
- Top 10%≤ 10%
- Good≤ 25%
- Medium≤ 50%
- Below avg> 50%
Speed
- Top 1%≥ 345 tok/s
- Top 5%≥ 237 tok/s
- Top 10%≥ 196 tok/s
- Good≥ 146 tok/s
- Medium≥ 90 tok/s
- Slow< 90 tok/s
Cost
- Freeopen weights · self-host
- Low< $1 / M out
- Moderate$1–5 / M out
- High≥ $5 / M out
Why it matters
A second large Chinese lab (after DeepSeek and Moonshot) shipping a reasoning-tuned 200B+ MoE under open weights tightens the gap between open and frontier closed models. Preview tag means scores will likely move on a stable release.
Core Capabilities
Long Documents
Handles entire codebases, books, and multi-doc RAG.
Generative
Produces images, video, audio, or other media.
Reasoning
Solves complex math, logic, and planning tasks.
Context Window
256k tokens
≈ 197+ pages
4k Chat 聊天
32k Long docs 长文档
128k Books 整本书
400k Multi-doc 多文档
1M Codebase 整个代码库
10M
256k
Availability
API
Not available
Product / App
Not available
Open Source
Released
Enterprise
—
Pricing Model
Free / self-host
Open weights — pay only for compute.
Self-host Capability / Performance
Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).
Lower 20% Upper 80% This model
Context / memory
Context window size · log-scaled
6.0
9.0
7.0
Lower 20% 20th percentile — 20% of models score below this This model Where the current model lands Upper 80% 80th percentile — only 20% of models score above this
Percentile boundaries are computed across every model in the tree that reports the underlying benchmark for each capability.
What it feels like
- Reasoning-style answers — explicit step-by-step traces rather than terse outputs, in the DeepSeek V3 / R1 lineage
- Frontier-class on benchmarks (AA Intelligence Index ~42), gap to GPT-5 / Opus 4.7 still meaningful but small among open-weights
Best use cases
- Self-host scenarios where you want a reasoning model with frontier benchmarks but no API dependency
- Math, coding, and graduate-physics-style problems (GPQA 86.7%, SciCode 41.2%)
- Long documents — 256K context window matches DeepSeek V3.2 and Kimi K2
Tools to try
Not ideal for
- Multimodal / vision tasks — text-only release
- Cost-sensitive serving without GPU capacity for a 295B/21B-active MoE
- Use cases that need a stable production SLA — this is a preview release