Hy3 Preview (Tencent Hunyuan 3, 295B MoE)

Deploy Tencent's Hy3 Preview (295B MoE, 21B active, 256K ctx) on Clore.ai — the first model from Tencent Hunyuan's rebuilt training stack, tuned for long-horizon reasoning and agentic coding

Status (April 2026): Hy3 Preview is the first public release from Tencent Hunyuan's rebuilt training infrastructure, published on April 13, 2026 and last updated April 23, 2026. Weights live at huggingface.co/tencent/Hy3-preview under the Tencent Hy Community License. Day-0 support landed in vLLM and SGLang.

Hy3 Preview is a 295B-parameter Mixture-of-Experts language model that activates only ~21B parameters per token (192 experts, top-8 routed). It targets two workloads where Tencent has been visibly catching up: long-horizon reasoning (FrontierScience-Olympiad, IMOAnswerBench, math-PhD exams) and agentic coding (SWE-bench Verified 74.4%, Terminal-Bench 2.0 54.4%, vendor-claimed). The 256K context window plus an MTP (Multi-Token Prediction) speculative-decoding layer make it practical for IDE-scale coding agents and document-heavy RAG.

For Clore.ai users, the headline number is 21B active. You don't need a full 8×H200 rack. A tensor-parallel deployment across 4×A100 80GB or 2×H100 80GB (BF16 with offload) is enough to serve it at usable throughput — frontier-class agentic coding for ~$10–20/day on the marketplace, with weights staying on your own box.

Key Specs

Property	Value
Total Parameters	295B (MoE)
Active Parameters	21B per forward pass
Experts	192 total, top-8 routed
Layers	80 transformer + 1 MTP
Attention	64 heads, GQA with 8 KV heads, head dim 128
Hidden Size	4096
Intermediate Size	13,312
Vocabulary	120,832
Context Window	256,000 tokens
Native Precision	BF16
License	Tencent Hy Community License
Release Date	April 13, 2026
Organization	Tencent Hunyuan
Primary Tooling	vLLM, SGLang, AngelSlim, LLaMA-Factory

Why Hy3 Preview?

• First on Tencent's rebuilt RL stack — Tencent rewrote its training infrastructure for this release; expect rapid iteration through 2026

• 21B active MoE — pay the inference cost of a ~21B dense model, not 295B

• 256K context — enough for full repos, long agent traces, or multi-document RAG in one shot

• MTP speculative layer — built-in multi-token prediction gives ~1.5–2× decode speedups on Hopper-class GPUs

• Two reasoning modes — reasoning_effort: "high" for chain-of-thought, "no_think" for fast direct answers

• Agentic-coding focus — explicitly tuned for SWE-bench-style multi-turn tool use and terminal agents

• Open-source-friendly license — Tencent Hy Community License is Apache-style for most uses; verify the LICENSE file for your case

---

Requirements

Still a 295B-class model. "21B active" describes inference compute, not the memory footprint. The full BF16 weights are ~590GB and must live in VRAM (or be offloaded). Plan for 8×H100/H200 if you want unconstrained throughput; 4×A100 80GB works with offload and shorter contexts.

Component	Minimum (Q4 GGUF, offload)	Recommended (BF16, TP)	Full BF16 (production)
GPU VRAM	~80GB + 256GB RAM offload	4× A100 80GB (320GB)	8× H100 80GB or 8× H20-3e
RAM	256GB	384GB	512GB
Disk	700GB NVMe	1TB NVMe	1.5TB NVMe
CUDA	12.4+	12.4+	12.6+
Driver	550+	550+	560+

Clore.ai pick: For most teams, 4× A100 80GB with BF16 tensor-parallel and --max-model-len 65536 is the sweet spot (~$10–16/day). If you need full 256K context with concurrent users, jump to 8× H100.

---

Option A — Ollama / GGUF (Quantized, community builds)

Heads-up: Hy3 Preview is brand new (April 13, 2026) and uses a custom MoE architecture. Community llama.cpp / GGUF support typically lands 2–4 weeks after release. If you need it today, use vLLM (Option B). Check huggingface.co/models?search=hy3-preview+gguf for community quants before pulling.

# Once a Q4_K_M build is published
docker exec ollama ollama pull hy3-preview:q4_K_M
docker exec ollama ollama run hy3-preview:q4_K_M

# Or with llama.cpp directly on a community GGUF
docker run --gpus all -it --rm -p 8080:8080 \
  -v $(pwd)/models:/models \
  ghcr.io/ggerganov/llama.cpp:server-cuda \
  -m /models/hy3-preview-q4_k_m.gguf \
  --n-gpu-layers 80 --ctx-size 32768 \
  --port 8080 --host 0.0.0.0

For pre-GGUF days, AngelSlim (Tencent's own quantization toolkit) can produce W4A16 / W8A8 weights directly from the BF16 checkpoint.

---

Option B — vLLM (Production API, recommended)

vLLM is Tencent's first-class serving target for Hy3 Preview. The MTP speculative layer is wired in via --speculative-config.method mtp.

version: "3.8"
services:
  vllm:
    image: vllm/vllm-openai:latest
    ports:
      - "8000:8000"
    volumes:
      - hf_cache:/root/.cache/huggingface
    command: >
      --model tencent/Hy3-preview
      --tensor-parallel-size 8
      --max-model-len 65536
      --gpu-memory-utilization 0.90
      --speculative-config.method mtp
      --speculative-config.num_speculative_tokens 1
      --tool-call-parser hy_v3
      --reasoning-parser hy_v3
      --enable-auto-tool-choice
      --served-model-name hy3-preview
      --trust-remote-code
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    shm_size: "16gb"

volumes:
  hf_cache:

# Test the API with high reasoning effort
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "hy3-preview",
    "messages": [
      {"role": "system", "content": "You are an expert software engineer."},
      {"role": "user", "content": "Refactor this Python function to use async/await and add proper error handling."}
    ],
    "max_tokens": 4096,
    "temperature": 0.9,
    "top_p": 1.0,
    "reasoning_effort": "high"
  }'

Reasoning modes. Set reasoning_effort: "high" to enable chain-of-thought traces (slower, much better on math/coding/agent tasks) or "no_think" for fast direct answers. The vendor-recommended sampling is temperature=0.9, top_p=1.0 — zero-temp sampling can break reasoning traces.

Tight on GPUs? Drop to --tensor-parallel-size 4 on 4× A100 80GB. Keep --max-model-len 32768 and add --enable-chunked-prefill to keep prefill latency reasonable.

---

Option C — SGLang

SGLang ships day-0 support and pairs the MTP layer with EAGLE speculative decoding for additional throughput on Hopper.

docker pull lmsysorg/sglang:latest

python3 -m sglang.launch_server \
  --model tencent/Hy3-preview \
  --tp 8 \
  --tool-call-parser hunyuan \
  --reasoning-parser hunyuan \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 1 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 2 \
  --mem-fraction-static 0.88 \
  --context-length 65536 \
  --served-model-name hy3-preview

Expect a 1.5–2× throughput boost on long agent loops compared to vanilla decode.

---

Clore.ai GPU Recommendations

Setup	VRAM	Expected Performance	Clore.ai Cost	Rent
4× A100 80GB	320GB	BF16 sharded, 64K ctx, ~15–25 tok/s	~$10–16/day	Rent A100 80GB
2× H100 80GB	160GB	BF16 with offload, smaller ctx, ~12–20 tok/s	~$12–18/day	Rent H100
8× H100 80GB	640GB	BF16 full, 256K ctx, 60+ tok/s with MTP	~$48–64/day	Rent H100
8× H200 141GB	1,128GB	BF16 full + max concurrency	~$64–96/day	Rent H200
1× RTX 5090	32GB	Q4 GGUF, RAM offload, single user	~$3.94/hr	Marketplace

Best value: 4× A100 80GB with BF16 tensor-parallel and a 64K context window. You get an open-weight 295B-class agentic coder for roughly the price of a Claude Pro subscription, and the weights never leave your rented box.

---

Use Cases

• Autonomous SWE agents — 74.4% SWE-bench Verified (vendor-claimed) and explicit tuning for long tool-call loops; pair with OpenHands, SWE-agent, or Aider

• Terminal-driven agents — 54.4% Terminal-Bench 2.0 puts it in the top tier for shell/CLI workflows

• Long-horizon reasoning — Olympiad-level math (IMOAnswerBench, FrontierScience-Olympiad) and PhD-grade STEM

• Codebase-scale RAG — 256K ctx fits a full mid-sized repo plus tests in a single prompt

• Search and browsing agents — BrowseComp / WideSearch tuning makes it a strong planner for multi-step web research

• Agent-of-agents — use Hy3 Preview as the planner and lighter open models (Qwen3.5, GLM-4.7 Flash) as workers

---

Benchmarks

Vendor-claimed — verify independently. All numbers below come from Tencent's April 13, 2026 model card. Independent reproductions (especially on SWE-bench Verified) are still rolling in. Treat them as upper bounds until LMSYS / OpenCompass confirms.

Benchmark	Hy3 Preview	GLM-5.1	DeepSeek R1	GPT-5.4
SWE-bench Verified	74.4%	~79%	~71%	~78%
Terminal-Bench 2.0	54.4%	—	—	—
GPQA Diamond	87.2%	—	~84%	~88%
SuperGPQA	51.6%	—	—	—
HLE	~30	—	—	—

Tencent also reports strong results on proprietary CL-bench / CL-bench-Life context-learning benchmarks and the Tsinghua Qiuzhen Math PhD exam (Spring 2026).

---

Troubleshooting

Issue	Solution
OutOfMemoryError on load	BF16 needs ~590GB total VRAM. Drop to 4×A100 with --max-model-len 32768 or use AngelSlim W4A16 quants.
Slow HuggingFace download	Use huggingface-cli download tencent/Hy3-preview --local-dir ./weights --resume-download. Expect 590GB+.
Tool calls silently dropped	Make sure --tool-call-parser hy_v3 (vLLM) or --tool-call-parser hunyuan (SGLang) is set, and --enable-auto-tool-choice is on.
Reasoning trace empty / wrong	Use temperature=0.9, top_p=1.0. Zero-temp greedy decoding breaks the chain-of-thought. Confirm reasoning_effort: "high".
MTP speculative decoding errors	Requires recent vLLM (post-April 2026 build). Run pip install -U vllm --pre or pin to a tag that lists mtp in release notes.
256K context OOMs	Start at --max-model-len 32768, enable --enable-chunked-prefill, raise gradually. Full 256K realistically needs 8× H200.
Custom architecture rejected	Always pass --trust-remote-code. Hy3 ships custom modeling code with the checkpoint.
Ollama / GGUF not available	Community quants typically arrive 2–4 weeks post-release. Use vLLM or AngelSlim in the meantime.

---

Next Steps

• Closest open-weight peer: GLM-5.1 — 744B / 40B-active MoE, MIT license, top SWE-bench Pro scores

• Multimodal alternative: Qwen3.5-Omni — text + audio + image + video, runs on a single RTX 4090

• Reasoning-only alternative: DeepSeek R1 — pure long-form reasoning specialist

• Rent the hardware: Rent A100 80GB on Clore.ai — 4× A100 80GB instances from ~$10/day

• Full marketplace: clore.ai/marketplace — H100, H200, A100, RTX 5090 from $0.50/day

Links

• Hy3 Preview on HuggingFace

• Hy3 Preview GitHub repo

• Tencent Hunyuan organization

• vLLM docs

• SGLang repo

• AngelSlim — Tencent's quantization toolkit