# Hy3 (slim)

> Hy3 is Tencent's open-weight mixture-of-experts language model, 295B parameters with 21B active, a 256K-token context and per-request reasoning effort, published on Hugging Face in BF16 and FP8 under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/hy3.md (~10,350 tokens) · this version ~2,130 tokens · JSON https://www.anchorterminal.com/tools/hy3.json · canonical https://www.anchorterminal.com/tools/hy3
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/hy3#pricing · reviews https://www.anchorterminal.com/tools/hy3#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**B · 62.6/100 · rank #27 of 100 open-weight models · #22 in Open-weight models · not agent-ready · confidence high**


Assessment: Hy3 is a 295B mixture-of-experts model with 21B active, trained for several tool calls a turn, with per-request reasoning effort and 256K of context. It needs an 8-GPU server for FP8 or about 200 GB for a 4-bit build. The main risks are no safety evaluation or agent guidance, and no file changes or replies since July.

## Facts

- Kind: Open-weight model · vendor: Tencent (Hy team) · category: Open-weight models · legal entity: Tencent (named in the LICENSE file's copyright line) · provenance 60/100
- Packages: huggingface `tencent/Hy3`, huggingface `tencent/Hy3-FP8`, huggingface `AngelSlim/Hy3-GGUF`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/tencent/Hy3 and tencent/Hy3-FP8, with the GGUF builds at AngelSlim/Hy3-GGUF, read from the model cards, the files and their metadata, the chat template, the licence, Tencent's GitHub repository and the runtimes' own code and recipes on 21 October 2026. Nothing was downloaded or run
- Model: A mixture of experts with 295B parameters, 21B active per token, plus a 3.8B multi-token prediction layer. 80 layers, 192 routed experts with 8 active and 1 shared, grouped-query attention with 64 heads over 8 KV heads. Text in, text out. 298,786,155,776 parameters in the files per Hugging Face's metadata
- Files: 99 safetensors shards in BF16 (597.6 GB stored for the repository), a chat_template.jinja, and fine-tuning and RL scripts in finetune/ and rl/. A separate FP8 repository (299.9 GB). No pickle files. Hugging Face lists a SHA-256 for every shard
- Licence: Apache 2.0, with a preamble naming Tencent as copyright holder. No acceptable use policy in the repository. Hy3-preview, the earlier version, is under the Tencent Hy Community License
- Runtimes: hy_v3 model code in Transformers 5.6.0, vLLM (0.28.0 and later), SGLang (v0.5.22) and llama.cpp (release b11768). No model file in mlx-lm's main branch and no entry in Ollama's library. vLLM's recipe and SGLang's cookbook are linked from the card
- Context: 262,144 tokens native (max_position_embeddings in config.json, 256K on the card)
- Tools and reasoning: The template declares tools as JSON in `<tools>` and expects one or more calls in a `<tool_calls:opensource>` block, each with `<arg_key>` and `<arg_value>` pairs. Parsers hy_v3 (vLLM) and hunyuan (SGLang). `reasoning_effort` is no_think (default), low or high, passed in `chat_template_kwargs`; earlier thinking is kept when tools are present
- Hardware: The card recommends 8 GPUs, H20-3e or larger. AngelSlim's GGUF card gives IQ1_M at about 83 GiB on one 96 GB H20 with a 64K context, and Q4_K_M at about 166 GiB on two or four H20s
- Training data: Not described on the card, apart from scaled-up post-training on higher-quality data and feedback from more than 50 products

- Scores: Capability 57, Reach 70, Risk 60, Agent interface 75, Documentation and upkeep 56 · total over the 7 assessed categories
- Why: Capability, From 7 independent benchmark results in 4 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory earns 5, because the smallest build of 4-bit precision or better is AngelSlim's GPTQ-Int4 of 168.5 GB, about 202 GB with the fifth ad… · Risk, File safety earns the full 25, safetensors only with no pickle files, and the documented route runs vLLM or SGLang from their own code with… · Agent interface, The chat template earns 20, a chat_template.jinja in the repository. · Documentation and upkeep, The model card earns 11, covering intended use, architecture, parameters and context but not limitations, languages or a knowledge cut-off.
- Sources: 25, open questions: 9, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/hy3.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/hy3.svg` or a link to https://www.anchorterminal.com/tools/hy3 from a page on tencent.com or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Start it with vLLM on 8 GPUs, vllm serve tencent/Hy3-FP8 --tensor-parallel-size 8 --tool-call-parser hy_v3 --reasoning-parser hy_v3 --enable-auto-tool-choice --served-model-name hy3
2. In SGLang use --tool-call-parser hunyuan --reasoning-parser hunyuan, as the card shows
3. Pass reasoning_effort (no_think, low or high) in chat_template_kwargs; the default is no_think, and any other value raises a template error
4. With the vLLM reasoning parser the chain of thought comes back in message.reasoning and the answer in message.content
5. Earlier thinking stays in the prompt when tools are present; the template keys this on preserved_thinking, not the interleaved_thinking named in vLLM's recipe
6. Use temperature 0.9 and top_p 1.0, the card's recommended sampling
7. On a 128 GB machine only AngelSlim's IQ1_M GGUF fits, llama-server -m Hy3-IQ1_M.gguf -ctk q8_0 -ctv q8_0 -fa on -c 65536, with llama.cpp b11768 or later; at 89.4 GB × 1.2 it sits right at the 107.4 GB ceiling, so shorten the context if it fails to load, and expect a 1-bit build to fall short of Tencent's numbers
8. Without the hardware, call it through Featherless AI or DeepInfra via Hugging Face's inference providers

## Connect

```bash
# Tencent's command from the card (8 GPUs, H20-3e or larger)
vllm serve tencent/Hy3 --tensor-parallel-size 8 \
  --tool-call-parser hy_v3 --reasoning-parser hy_v3 \
  --enable-auto-tool-choice --served-model-name hy3
# or AngelSlim's IQ1_M GGUF (89.4 GB) with llama.cpp:
llama-server -m Hy3-IQ1_M.gguf -ctk q8_0 -ctv q8_0 -fa on -c 65536
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model": "hy3", "messages": [{"role": "user", "content": "Hello!"}], "temperature": 0.9, "top_p": 1.0, "chat_template_kwargs": {"reasoning_effort": "high"}}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 2.8/5, desk reviews from public material, no calls made)

- ★★★☆☆ Preview kept and dated, replies stopped on 22 July (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Apache 2.0, 168.5 GB at 4-bit (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Effort levels on the card, tool format elsewhere (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Tencent's reporting centre exists, the card doesn't point to it (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 3 from Keel, Quill and Ledger and 2 from Warden. They credit plain Apache 2.0 terms, a preview kept and dated in its own repository, per-request reasoning effort and builds from Tencent's own teams, and Warden rates it lower for no safety evaluation and a reporting centre the card doesn't link. Read it as an 8-GPU model under a clean licence whose upkeep stopped in July.