1. Summary
Agents-A1-4B is a 4B-parameter dense agent model released by Shanghai Artificial Intelligence Laboratory (InternScience) on July 14, 2026. It possesses heterogeneous agent capabilities spanning five domains: long-horizon search, engineering and scientific research, instruction following, and tool calling. The model records scores significantly higher than Qwen3.5-4B, a model of similar size, across multiple benchmarks. Notably, it achieves the highest scores among models of its size on XBench-DS-2510 (90.0), GAIA (95.1), and IFEval (94.8). On benchmarks such as BrowseComp and FrontierScience-Research, it matches or exceeds the performance of larger MoE models like Nex-N2-mini and Qwen3.6. With a 262K context window, support for SGLang and vLLM, and an Apache-2.0 license, it is suitable for building local AI assistants and integrating agentic workflows.
2. Benchmark Results
The following figures are from the "Performance" section of the Hugging Face model card. 🥇 indicates the highest score in the table.
| Benchmark | Qwen3.5-4B | Agents-A1-4B | Qwen3.5 | Qwen3.6 | Nex-N2-mini | Agents-A1 (35B-A3B) |
|---|---|---|---|---|---|---|
| Long-horizon Search | ||||||
| BrowseComp | 47.2 | 66.8 | 61.0 | 67.9 | 74.1 | 🥇 75.5 |
| XBench-DS-2510 | 73.0 | 🥇 90.0 | 77.0 | 71.0 | 82.0 | 86.0 |
| Seal0 | 31.5 | 45.8 | 41.4 | 38.7 | 49.6 | 🥇 56.4 |
| GAIA | 58.3 | 95.1 | 59.8 | 78.6 | 82.5 | 🥇 96.0 |
| Engineering & Research Tasks | ||||||
| SciCode | 16.1 | 29.6 | 37.7 | 35.8 | 29.9 | 🥇 44.3 |
| MLE-Lite | 7.6 | 22.7 | 24.2 | 34.9 | 34.9 | 🥇 43.9 |
| LiveCodeBench-V6 | 55.8 | 59.6 | 76.2 | 🥇 78.1 | 59.1 | 76.2 |
| FrontierScience-Research | 1.7 | 33.3 | 2.5 | 2.9 | 5.0 | 🥇 40.0 |
| Instruction Following | ||||||
| IFBench | 59.2 | 69.1 | 70.2 | 64.4 | 54.1 | 🥇 80.6 |
| LongBench-v2 | 50.0 | 52.1 | 59.0 | 57.7 | 59.6 | 🥇 60.2 |
| IFEval | 89.8 | 🥇 94.8 | 91.9 | 91.3 | 88.4 | 🥇 94.8 |
| General & Scientific Agents | ||||||
| τ²-Bench | 79.9 | 78.2 | 🥇 81.2 | 79.0 | 74.5 | 79.8 |
| VitaBench | 22.0 | 🥇 40.3 | 31.9 | 35.6 | 23.0 | 38.8 |
| MatTools | 10.9 | 🥇 49.3 | 21.0 | 15.9 | 34.1 | 47.1 |
Source: Hugging Face model card, "Performance" section.
3. Official Announcements
- 2026.7.14: Released the 4B model (Agents-A1-4B).
- 2026.7.8: Announced that a 4B model would be released within days, responding to community requests.
- 2026.7.2: Released quantized model variants of the Agents-A1 series. Multiple scale quantized builds for Mac provided by mlx-community are also available.
- 2026.6.26: Open-sourced the flagship Agents-A1 35B-A3B model, evaluation code for selected domains, and the technical report (arXiv:2606.30616).
Source: Hugging Face model card, "🔥 News" section.
4. Real-world Performance (Community Reception)
Sufficient information has not yet been gathered.
5. Recommended Parameters
The generation parameters recommended in the "Recommended Sampling Parameters & System Prompt" section of the Hugging Face model card are as follows:
| Parameter | Recommended Value |
|---|---|
| temperature | 0.85 |
| top_p | 0.95 |
| top_k | 20 |
| min_p | 0.0 |
| presence_penalty | 1.1 |
| repetition_penalty | 1.0 |
The model card also includes a recommended system prompt ("You are Intern-A1, a deep research assistant developed by InternAgent Team, Shanghai Artificial Intelligence Laboratory. …"). This configuration is intended to stabilize multi-turn behavior.
6. Sources
- Hugging Face model card (Official)
2. Bunrin Bench (Bunrin AI Lab's Proprietary Test)
No data (scheduled for implementation). Refer to About Bunrin Bench for methodology.