Vireqo-27T-260818
Vireqo-27T-260818 is the bounded-Thinking product built on the unchanged Vireqo-27B-260816 Q1 language payload. The T product identity is defined by the tested Think512-Concise LM Studio configuration, not by a new tensor conversion.
中文摘要：这是独立于 27B 的 Thinking 产品线。它复用同一份外置盘主权重，但通过独立目录、名称、preset、模型卡和验收报告发布。原 Vireqo-27B-260816 继续保持 Thinking Off 定位。
Bundle
| Item | Value |
|---|---|
| Main file | Vireqo-27T-260818.gguf |
| Physical source | unchanged Vireqo-27B-260816.gguf |
| Size | 4,876,198,496 bytes / 4.5413 GiB |
| SHA-256 | 54511bc32a461c6558d0c74278f4e75357cc9dc9fc2b39fbd5b52940ddc031d5 |
| Preset | thinking-preset.json |
| New physical main-weight copy | no |
The local GGUF filename is a symbolic-link alias. Its internal general.name remains Vireqo-27B-260816; the T product name is supplied by the release directory and LM Studio model key.
Think512-Concise
| Setting | Value |
|---|---|
| Think | On |
| Reasoning Budget | 512 |
| Maximum response length | 768 |
| Temperature | 0 |
| Top-p | 1 |
| Repeat penalty | 1.08 |
| Context | 2048 |
| Parallel | 1 |
Load-time Reasoning Budget Message:
思考预算已到。复核已有结论后，只输出最终答案；禁止重复或展示思考过程。
Validation
LM Studio SDK / engine-protocol validation passed 3/3:
- France capital →
法国的首都是巴黎。 17×19→323- chicken/rabbit system →
23chickens and12rabbits
All probes produced nonempty reasoning, clean non-reasoning final content, normal EOS, and no internal synthetic marker leak. See thinking-validation.md.
Budget ablation showed that 64/192/256 can cut through the core calculation and produce incorrect answers; 512 plus the strict budget message is the first accepted preset.
Use
In LM Studio, load Vireqo-27T-260818, open Advanced load settings, set the budget message above, then enable Think and Reasoning Budget 512 in Chat. See LM-STUDIO-使用指南.md.
Limitations
- Unrestricted Thinking is not accepted and may return empty visible content.
- The preset is optimized for concise finals, not long explanatory answers.
- This is not a retrained reasoning model or a standard reasoning benchmark claim.
- Complex math, code, or planning may require a future larger-budget product.
- Do not use for high-stakes decisions.
See bundle-provenance.json, thinking-preset.json, and TECHNICAL_README.md.