vLLM
The [Model] Support Hy3 preview pull request was merged on April 23, 2026.
SourceRetrieved 2026-08-12
Model reference
Hy3
This reference covers the open mixture reasoning and agent checkpoint with a quantized companion. It separates text-backed benchmark claims from the image-only appendix and keeps provider pricing distinct.
Hy3 is the vendor reference for tencent/Hy3, as of 2026-08-12. Model scale: 295B total; 21B activated; 3.8B MTP; Context length: 256K, as of 2026-08-12. RunInfra has not measured this model; every figure below belongs to its named source, cited and dated.
Cited specifications
| Group | Fact | Source-cited display value | Source and date |
|---|---|---|---|
| Identity | Release identity | tencent/Hy3 | SourceRetrieved 2026-08-12 |
| Identity | Vendor description | Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters and 3.8B MTP layer parameters, developed by the Tencent Hy Team. | SourceRetrieved 2026-08-12 |
| Identity | Lineage | Following the Hy3 Preview launch in late April, we gathered feedback from 50+ products and scaled up post-training with higher quality data. | SourceRetrieved 2026-08-12 |
| Identity | OpenRouter release date | Jul 6, 2026 | SourceRetrieved 2026-08-12 |
| Architecture | Model scale | 295B total; 21B activated; 3.8B MTP | SourceRetrieved 2026-08-12 |
| Architecture | Layer configuration | 80 layers excluding MTP; 1 MTP layer | SourceRetrieved 2026-08-12 |
| Architecture | Expert routing | 192 experts, top-8 activated | SourceRetrieved 2026-08-12 |
| Architecture | Attention | 64 (GQA, 8 KV heads, head dim 128) | SourceRetrieved 2026-08-12 |
| Architecture | Hidden and intermediate sizes | hidden 4096; intermediate 13312 | SourceRetrieved 2026-08-12 |
| Architecture | Vocabulary | 120832 | SourceRetrieved 2026-08-12 |
| Architecture | Precision | BF16 | SourceRetrieved 2026-08-12 |
| Architecture | Vendor architecture framing | Built on a hybrid fast-and-slow-thinking Mixture-of-Experts (MoE) architecture | SourceRetrieved 2026-08-12 |
| Context | Context length | 256K | SourceRetrieved 2026-08-12 |
| Modalities | Pipeline tag | text generation | SourceRetrieved 2026-08-12 |
| License | Weights license | Hy3 is released under the Apache License 2.0. | SourceRetrieved 2026-08-12 |
| License | License file | Apache License, Version 2.0; Copyright (C) 2026 Tencent. All rights reserved. | SourceRetrieved 2026-08-12 |
| Pricing | OpenRouter listing | $0.1288 per 1M input tokens / $0.5336 per 1M output tokens | SourceRetrieved 2026-08-12 |
| Availability | Repository access | Ungated repository. | SourceRetrieved 2026-08-12 |
| Availability | Weight distribution | We open-source Hy3 and Hy3-FP8 model weights on Hugging Face, ModelScope, GitCode, and CNB. | SourceRetrieved 2026-08-12 |
| Availability | FP8 companion | tencent/Hy3-FP8; FP8 quantized instruct model | SourceRetrieved 2026-08-12 |
| Availability | WorkBuddy access | available free of charge to users worldwide until 31 August 2026 (Pacific Time) | SourceRetrieved 2026-08-12 |
| Availability | vLLM recipe floor | vLLM 0.26.0+ | SourceRetrieved 2026-08-12 |
| Availability | Vendor serving recommendation | For production serving, we recommend using vLLM or SGLang, both of which provide dedicated recipes for Hy3 | SourceRetrieved 2026-08-12 |
| Availability | Main weights | tencent/Hy3 | SourceRetrieved 2026-08-12 |
| Availability | Quantized instruct weights | tencent/Hy3-FP8 | SourceRetrieved 2026-08-12 |
What the vendor says is new
"Today, we introduce Hy3, which outperforms similar-size models and rivals flagship open-source models with 2-5x parameters."
SourceRetrieved 2026-08-12
"Hy3 recorded more than 68 times as many API calls as the previous-generation model and ranked first globally on OpenRouter's global LLM usage leaderboard within one week of launch."
SourceRetrieved 2026-08-12
Vendor-claimed benchmarks
These results are vendor-claimed, not independently measured by RunInfra.
Serving concepts
What RunInfra measures when we measure it
RunInfra has not measured this model yet. When measurement is published, the record will state throughput, latency, memory, quality, serving conditions, and reproducible evidence.
Questions about this model reference
What is Hy3?
As of 2026-08-12, release identity: tencent/Hy3.
Where are the Hy3 weights?
As of 2026-08-12, repository access: Ungated repository.
Can I run Hy3 myself?
As of 2026-08-12, repository access: Ungated repository.
RunInfraby RightNow© 2026 RunInfra. All rights reserved.
Ask AI about RunInfra
