Qwen3.8 27B is an open-weight dense vision-language model from Qwen.
Hy3
Hy3 is a 295B-parameter Mixture-of-Experts model from Tencent (21B active, 192 experts with top-8 routing) built for reasoning, agentic work
SharkScore
#79 in Models
Why it scores 43.3
Hy3 has a SharkScore of 43.3/100, ranked #79 of 607 in Models (updated 2026-11-19).
Each part compares Hy3 with other listings in Models. Evidence covers 87% of the score's weight.
- Adoption
- 33 · 30%
- User reviews
- no data · 25%
- Maintenance
- 28 · 20%
- Trust & transparency
- 80 · 15%
- Momentum
- no data · 10%
Paying never changes this score. How ranking works · method v2026-11-18.2
RankShark review of Hy3
Automated evaluation by RankShark · 18 Nov 2026 · method v1.0 · How we review
We were not paid for this review.
As of 2026-11-18, Hy3 is Tencent’s text-generation MoE model; the OpenRouter listing shows 262,144 tokens of context and prices of $0.132 per million input tokens and $0.528 per million output tokens. [E1, E2]
Hy3 is a Tencent 295B-parameter mixture-of-experts text-generation model with 21B active parameters. Tencent publishes model weights and instructions for serving it with vLLM or SGLang. [E1]
Best for
- Developers evaluating a model for reasoning, coding, or agent workflows. · model card
- Teams that want to self-host a model using vLLM or SGLang. · model card
Not for
- Deployments that cannot allocate multiple high-memory GPUs for self-hosting; Tencent recommends 8 GPUs and H20-3e or other GPUs with larger memory capacity. · model card
- Applications requiring image, audio, or other non-text input or output; the OpenRouter listing specifies text input and text output. · OpenRouter
Key features
- Mixture-of-experts architecture. The vendor says Hy3 has 295B total parameters, 21B active parameters, and 192 experts with 8 activated. · model card
- Long context. The vendor lists a 256K context length; the OpenRouter listing reports 262,144 tokens. · model card, OpenRouter
- Configurable reasoning. The vendor says the reasoning_effort setting can be set to no_think, low, or high; the OpenRouter listing also reports supported reasoning-effort settings. · model card, OpenRouter
- Tool calling and output formats. The vendor says it improved tool-call error recovery and support across tool configurations and output constraints. · model card
- Self-hosting guidance. The vendor provides deployment recipes for vLLM and SGLang and recommends 8 GPUs for serving. · model card
- Open weights. The vendor says Hy3 and Hy3-FP8 weights are available on Hugging Face, ModelScope, GitCode, and CNB. · model card
Pros
- The OpenRouter listing reports a context window of 262,144 tokens. · OpenRouter
- The vendor provides model weights and deployment instructions for vLLM and SGLang. · model card
- The vendor provides a finetuning pipeline and says Hy3 supports GRPO reinforcement-learning training with verl. · model card
Cons
- Tencent’s self-hosting guidance recommends 8 GPUs and GPUs with larger memory capacity, which may limit deployment options. · model card
- The OpenRouter listing identifies text as the model’s input and output modality. · OpenRouter
- No free plan or free trial stated on the pages we checked.
- Whether customer data is used for model training is not stated on the pages we checked.
Pricing
| Plan | Price | Limits |
|---|---|---|
| OpenRouter inference | $0.132 per million input tokens; $0.528 per million output tokens | OpenRouter lists a 262,144-token context window and a maximum of 128,000 completion tokens. |
Checked 18 Nov 2026
- Free plan
- Not stated (checked 18 Nov 2026)
- Free trial
- Not stated (checked 18 Nov 2026)
- Trains on customer data
- Not stated (checked 18 Nov 2026)
- Data retention
- Not stated (checked 18 Nov 2026)
- Data regions
- Not stated (checked 18 Nov 2026)
- Data processing agreement
- Not stated (checked 18 Nov 2026)
- AI models used
- Hy3; Hy3-FP8 (FP8-quantized instruct model) · model card
- API
- The vendor documents an OpenAI-compatible API for self-hosted deployments. OpenRouter lists chat completions, completions, messages, and responses endpoints. · model card, OpenRouter
- SDKs
- The vendor’s quickstart shows the OpenAI Python client. · model card
- Webhooks
- Not stated (checked 18 Nov 2026)
- Platforms
- Not stated (checked 18 Nov 2026)
- Security claims
- No security certification stated (checked 18 Nov 2026)
What we measured
- Context window (OpenRouter listing) · now
- 262,144 tokens
- Price per million tokens (OpenRouter listing) · now
- $0.132 in, $0.528 out
- Package downloads · 30 days
- 987,210
How we tested · sources
- Hourly uptime and response-time checks of the product's URL.
- Repository, release and download data from GitHub, npm, PyPI and Hugging Face where they exist.
- We did not sign up, pay, or use the product as a customer.
- model card: https://huggingface.co/tencent/Hy3 · fetched 18 Nov 2026
- OpenRouter: https://openrouter.ai/tencent/hy3 · fetched 18 Nov 2026
- 18 Nov 2026: Plans or prices changed on the pricing page.
- 18 Nov 2026: Plans or prices changed on the pricing page.
Something wrong? Report an error with a public link that shows it; corrections re-run the review.
Top alternatives to Hy3
See all- 67.1ModelOpen source
- 65.5
Qwen3.6-35B-A3B is an open-weight multimodal model from Alibaba Cloud with 35 billion total parameters and 3 billion active parameters per t
ModelOpen source - 64.7
Qwen3.8 Flash is a multimodal reasoning model from Alibaba.
ModelOpen source - 64.3
Qwen3.6 27B is a dense 27-billion-parameter language model from the Qwen Team at Alibaba, released in April 2026.
ModelOpen source