Hy3 preview
Language model
Hy3 preview
Tencent
- Released
- April 20, 2026
- Data date
- September 26, 2026
- Input
- Text
- Output
- Text
Hy3 preview was released as an open-weight preview for practical agent work. Tencent uses a Mixture-of-Experts model with 295 billion total parameters, 21 billion active per token, and a 262,144-token context. The model supports no_think for direct responses and high for harder reasoning tasks.
The weights run locally with vLLM or SGLang. Tencent Cloud has routed hy3-preview calls in its Token Plan to Hy3 since July 24, 2026. The published preview remains a distinct checkpoint for inspection and local use, while the API alias points to Hy3 there.
Specifications and access
| Specification | Value and source |
|---|---|
| Weights | Open weightsSource |
| Access | APISource |
| Parameters | 295 billionSource |
| Active parameters | 21 billionSource |
| Context window | 256,000 tokensSource |
| Maximum output | 64,000 tokensSource |
| License | Tencent Hunyuan CommunitySource |
| Inputs | TextSource |
| Outputs | TextSource |
| Reasoning | YesSource |
| Tool calling | YesSource |
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
SWE-bench Verified (Show measurement, test conditions, and source)
- Source value
- 74.4
- Score
- 74.4
- Metric
- resolved rate
- Unit
- %
- Direction
- Higher is better
- Source type
- official-leaderboard
- Evaluator
- SWE-bench Verified official leaderboard
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Official benchmark leaderboard. Exact Tencent Hy3-preview model result on SWE-bench Verified publisher leaderboard; resolved-rate metric. Run date, agent harness, and inference settings are not stated. Benchmark prompt, sampling/decoding settings, reasoning effort, quantization, and task run date are not stated in the cited result; none are inferred.
GPQA Diamond (Show measurement, test conditions, and source)
- Source value
- 87.2
- Score
- 87.2
- Metric
- score
- Unit
- No unit provided
- Direction
- Higher is better
- Source type
- official-leaderboard
- Evaluator
- GPQA official benchmark result surfaced on Tencent model card
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Official benchmark leaderboard. Exact Hy3-preview leaderboard widget for GPQA Diamond; prompt, sampling, effort, quantization, unit, and run date are not stated. Benchmark prompt, sampling/decoding settings, reasoning effort, quantization, and task run date are not stated in the cited result; none are inferred.
- Context
- The source presents a raw score without an explicit unit or scale; the published value is retained without percentage conversion.
Terminal-Bench 2.0 (Show measurement, test conditions, and source)
- Source value
- 54.4
- Score
- 54.4
- Metric
- score
- Unit
- No unit provided
- Direction
- Higher is better
- Source type
- official-leaderboard
- Evaluator
- Terminal-Bench 2.0 official leaderboard
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Official benchmark leaderboard. Exact Hy3-preview leaderboard widget for Terminal-Bench 2.0; harness/agent setup, effort, quantization, unit, and run date are not stated. Benchmark prompt, sampling/decoding settings, reasoning effort, quantization, and task run date are not stated in the cited result; none are inferred.
- Context
- The source presents a raw score without an explicit unit or scale; the published value is retained without percentage conversion.
Humanity's Last Exam (Show measurement, test conditions, and source)
- Source value
- 30
- Score
- 30
- Metric
- score
- Unit
- No unit provided
- Direction
- Higher is better
- Source type
- official-leaderboard
- Evaluator
- Hugging Face HLE evaluation result widget on Tencent model card
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Official benchmark leaderboard. Exact Hy3-preview model-card evaluation result; the displayed value includes an asterisk, whose interpretation is not explained on the card. Tools/setup, task date, and unit are not stated. Benchmark prompt, sampling/decoding settings, reasoning effort, quantization, and task run date are not stated in the cited result; none are inferred.
- Context
- The publisher card displays an asterisk beside this score and does not explain it; retained literally as a raw value with unknown unit.
Our tests
SVG short test
One SVG prompt and fixed output settings. Captures are shown only with 0 reported reasoning tokens or documented lack of reasoning controls. Identical test conditions do not make the models' internal compute identical.
Earlier tests
Pelican on a bicycle
A community classic for free-form SVG drawing.
Hy3 preview
The response reached the token limit. No complete recording is available.
Reasoning (requested): high
Current attempt
- Run date
- Oct 5, 2026
- API provider
- openrouter
- Requested API model
- tencent/hy3-preview
- API model reported in response
- tencent/hy3-preview
- Reasoning setting
- high
- Pinned endpoint
- GMICloud | tencent/hy3-preview-20260421
- Temperature
- Provider default
- Token limit including reasoning
- 8,192 tokens
- Test protocol
- community-visual-tests-v1
- Technical checks
- Failed
- Response time
- 62.3 s
Not yet visually reviewed. A successful recording does not confirm correct physics or full compliance with the prompt.
Earlier attempts (2)
- Oct 2, 2026API request failedReasoning setting: high
- Oct 1, 2026API request failedReasoning setting: high
Prompt and test conditions
Generate an SVG of a pelican riding a bicycle
Token limit including reasoning: 8,192 tokens
Task origin (Simon Willison)Popularity
35.8 / 100