Hy3 Preview
LLM Provider / Model · Agent-led tasks in development environments, CodeBuddy/WorkBuddy workflows, low-latency production coding (54% TTFT reduction), open-weight frontier alternative
At a glance
| Input price | $0.30 |
| Output price | $1.20 |
| Cache price | $0.03 |
| Price tier | Budget |
| Context window | 256K |
| Max output | 32K |
| Context tier | 128K-500K |
| Speed tier | Standard |
| Latency | medium |
| Knowledge cutoff | Feb 2026 |
| Modality | Text-only |
| Model ID | hy3-preview |
| Provider | Tencent |
| HumanEval | 90% |
| MMLU | 86% |
| SWE-Bench | 74% |
| Benchmark | SWE-bench Verified: high |
| Released | Apr 2026 |
| Hosting | Open-weights |
| Capabilities | Tool use, Streaming, Structured output, Agentic coding, Multi-Token Prediction (3.8B layer), 295B MoE (21B active), Function calling, Multi-step reasoning, Self-hostable |
What Hy3 Preview does
Tool use, Streaming, Structured output, Agentic coding, Multi-Token Prediction (3.8B layer), 295B MoE (21B active), Function calling, Multi-step reasoning, Self-hostable
Best for
Agent-led tasks in development environments, CodeBuddy/WorkBuddy workflows, low-latency production coding (54% TTFT reduction), open-weight frontier alternative
Works well with
Agent / Orchestration
Conflicts & caveats
- Trae's built-in model menu may not include Hy3 Preview — Trae primarily ships with Claude/GPT/Gemini/DeepSeek.
- Google Antigravity is built around Gemini 3 Pro; Claude and recent GPT-5 models also work, but Hy3 Preview is not in the supported model list.
- Kiro's spec mode is tuned for Claude models — Hy3 Preview may work but spec-driven flows are best with Claude.
- Cursor's built-in model menu may not include Hy3 Preview — verify it's available in your Cursor model picker.
- OpenAI Codex (VS Code) is tied to OpenAI's GPT-5.1 Codex family — Hy3 Preview is not selectable inside this integration.
- Gemini CLI is locked to Google Gemini models — your LLM is Hy3 Preview.
- Privacy conflict: Self-hosted Hy3 Preview sends code to cloud @codebase (Cursor). Use local context (Continue indexing, ChromaDB, LanceDB, pgvector, Vespa self-hosted) for true privacy.
- Claude Code only works with Claude models — your LLM is Hy3 Preview.
- Devin uses Cognition's proprietary model internally — your selected Hy3 Preview will be ignored.
- OpenAI Codex (Cloud) runs GPT-5.1 Codex internally — your selected Hy3 Preview will be ignored.
- Jules runs Gemini models internally — your selected Hy3 Preview will be ignored.
- The Claude Code SDK is tuned for Claude models — Hy3 Preview may work via the model adapter but loses Claude-specific features (thinking, prompt caching, sub-agents).
- Lovable uses a fixed internal model — your selected Hy3 Preview will be ignored.
- Replit Agent 3 uses its own model selection internally — your selected Hy3 Preview will be ignored.
Build a full stack around Hy3 Preview — Flowpicker shows compatibility warnings before you commit.
Open the stack planner →