Links & Resources
Tencent open-sourced Hy3 preview: a Mixture-of-Experts model with 295B total parameters but only 21B active. It runs at the cost of a 21B model while trading blows with DeepSeek-V3, Kimi-K2, and proprietary flagships several times larger.
The headline trick: it blends fast and slow thinking and decides on its own when a question needs deep reasoning - cutting wasted tokens instead of burning them on trivia.
The Specs
- Architecture: MoE, top-8 of 192 experts
- Params: 295B total / 21B active
- Context: 256K tokens
- Precision: BF16
- License: Tencent Hy Community License
- Weights: Hugging Face, ModelScope, GitCode
For comparison: Kimi-K2 is 1043B total (32B active), DeepSeek-V3 is 671B (37B active). Hy3 is smaller and cheaper per token.
Fast + Slow Thinking
Most reasoning models have one gear: think hard about everything. Hy3 exposes a reasoning_effort control with three modes - no_think, low, high - and defaults to not reasoning unless the task needs it.
Tencent reports from its own CodeBuddy and WorkBuddy deployments:
- Time To First Token down 54%
- End-to-end response time down 47%
- 40% better inference efficiency at comparable cost
Vendor-reported, not independently audited - but the direction matches what a hybrid-reasoning design should buy you.
Where It Wins (and Loses)
Wins the reasoning-per-cost fight:
- MATH (4-shot): 76.28 vs Kimi-K2 71.20, DeepSeek-V3 59.37
- GSM8K: 95.37 (top)
- LiveCodeBench-v6: 34.86 (top)
- SWE-bench Verified: 74.4
- Terminal-Bench 2.0: 54.4
- GPQA Diamond: 87.2
Loses on memorized-knowledge trivia (MMLU, SimpleQA, Chinese-SimpleQA) - the bigger models still know more facts. Honest read: trades recall for reasoning.
Plug It Into Your Tools
Hy3 serves an OpenAI-compatible API - anything with a custom base URL works (Claude Code, Cursor, Cline). Official integrations: OpenClaw, OpenCode, KiloCode.
Easiest way to try it - the free OpenRouter route:
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "tencent/hy3-preview:free",
"messages": [{"role": "user", "content": "Refactor this function and add tests."}]
}'
Pricing
Free two-week launch window on OpenRouter (hy3-preview:free). Steady use via Tencent Cloud TokenHub:
- Input: ~$0.18 / 1M tokens
- Cached input: ~$0.06 / 1M tokens
- Output: ~$0.59 / 1M tokens
Self-hosting is $0 in license fees - you just pay for the 8-GPU rig to serve it (H20-3e or larger recommended).
The Takeaway
A 21B-active open-weights model matching flagship SWE-bench numbers, with a built-in switch to skip reasoning on easy questions. If you're paying per token or per second, that math gets interesting fast.