| --- |
| license: apache-2.0 |
| language: |
| - en |
| base_model: Qwen/Qwen3.5-0.8B-Base |
| pipeline_tag: text-generation |
| tags: |
| - reasoning |
| - cot |
| - math |
| - gpqa |
| - gsm8k |
| - quasar-1 |
| --- |
|  |
|
|
| # 🌌 Quasar-1-0.8B |
|
|
| **Quasar-1-0.8B** is a compact reasoning-oriented language model built upon **Qwen/Qwen3.5-0.8B-Base**. The model is designed to maximize multi-step mathematical reasoning and stable zero-shot execution while remaining lightweight enough for local inference on highly constrained hardware. |
|
|
| ## Chat with the model: |
|
|
| here is the 🤗 space to chat with the model [Ion-Chat](https://huggingface.co/spaces/Axiom-AI/Ion-Chat) |
|
|
| ## 📌 Highlights |
| * 🚀 **Sub-1B parameter model** |
| * 🧮 Strong zero-shot mathematical reasoning |
| * 📉 Efficient local execution |
| * 🔍 Fully reproducible evaluation artifacts |
| * 📂 Raw benchmark logs included |
| * ⚖️ **Apache-2.0 licensed** |
|
|
| --- |
|
|
| # 📊 Model Overview |
|
|
| | Property | Value | |
| | :--- | :--- | |
| | **Base Model** | [Qwen/Qwen3.5-0.8B-Base](https://huggingface.co) | |
| | **Parameters** | ~0.8B | |
| | **Architecture** | Transformer | |
| | **License** | Apache-2.0 | |
| | **Intended Use** | Reasoning, mathematics, local inference | |
| | **Precision Tested** | `float16` | |
|
|
| --- |
|
|
| # 📈 Evaluation |
|
|
| All results below correspond to clean baseline runs performed with [lm-evaluation-harness](https://github.com). No scores were manually edited or cherry-picked. |
|
|
| ## Zero-Shot Performance |
|
|
| | Model | GPQA (0-shot) | GSM8K (0-shot) | |
| | :--- | ---: | ---: | |
| | **Quasar-1-0.8B** | **19.7%** | **39.8%** | |
| | OrionLLM/GRM-2.5-Air | 12.5% | — | |
| | Qwen/Qwen3.5-0.8B-Base | 11.9% | ~15–20% | |
| | Random Guessing Baseline | 25.0% | ~0% | |
|
|
| --- |
|
|
| # 🔍 Observations |
|
|
| * **Stable Zero-Shot Execution:** Quasar-1-0.8B consistently produces structured outputs under default evaluation settings, allowing direct usage with standard `lm-evaluation-harness` configurations. |
| * **Mathematical Capability:** Despite operating at sub-1B scale, the model demonstrates strong performance on [GSM8K](https://huggingface.co), suggesting useful multi-step arithmetic and reasoning capabilities. |
| * **Edge Efficiency:** The model is intended for local execution on low-resource hardware, enabling experimentation without requiring modern accelerators. |
|
|
| --- |
|
|
| # 💻 Hardware Used |
|
|
| The reported results were obtained on standard consumer hardware: |
|
|
| | Component | Specification | |
| | :--- | :--- | |
| | **OS** | Windows 10 Pro 64-bit | |
| | **CPU** | Intel Pentium G4560 @ 3.50 GHz | |
| | **GPU** | NVIDIA GTX 1050 Ti | |
| | **Framework** | PyTorch 2.6.0 | |
| | **CUDA** | 12.4 | |
| | **Transformers** | 5.10.2 | |
| | **lm-eval** | 0.4.13.dev0 | |
| | **Precision** | `float16` | |
|
|
| --- |
|
|
| # 📂 Reproducibility Artifacts |
|
|
| Included evaluation files: |
| * `gpqa_results_2026-06-05T17-38-23.426157.json` |
| * `results_2026-06-09T15-53-28.196995.json` |
|
|
| These files contain: |
| * Runtime environment information |
| * Framework versions |
| * Evaluation settings |
| * Benchmark metrics |
| * Full telemetry metadata |
|
|
| --- |
|
|
| # 🎯 Intended Uses |
|
|
| Suitable for: |
| * Mathematical reasoning |
| * Educational experiments |
| * Lightweight local inference |
| * Edge devices |
| * Research on small language models |
|
|
| # 🛑 Limitations |
|
|
| Quasar-1-0.8B remains a compact model and should not be expected to match larger frontier systems on broad knowledge or complex long-horizon reasoning tasks. Benchmark scores represent only the evaluated tasks and should not be interpreted as comprehensive measurements of intelligence or general capability. |
|
|
| --- |
|
|
| # 📜 Citation |
|
|
| ```bibtex |
| @misc{quasar1, |
| title={Quasar-1-0.8B}, |
| year={2026}, |
| license={Apache-2.0}, |
| base_model={Qwen/Qwen3.5-0.8B-Base} |
| } |
| ``` |
|
|
| # ⚖️ License |
|
|
| Released under the **Apache-2.0 License**. |
|
|
