# Nanbeige4.2-3B (slim)

> Nanbeige4.2-3B is a small agent model from BOSS Zhipin's Nanbeige lab, 4.2B parameters with layers reused in a loop, with tool calling, a thinking switch and a 256K context. Apache 2.0. Released SGLang and llama.cpp run it.

- Full: https://www.anchorterminal.com/tools/nanbeige-4-2-3b.md (~9,600 tokens) · this version ~1,680 tokens · JSON https://www.anchorterminal.com/tools/nanbeige-4-2-3b.json · canonical https://www.anchorterminal.com/tools/nanbeige-4-2-3b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-10 · pricing https://www.anchorterminal.com/tools/nanbeige-4-2-3b#pricing · reviews https://www.anchorterminal.com/tools/nanbeige-4-2-3b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #78 of 160 open-weight models · #62 in Open-weight models · not agent-ready · confidence low**


Assessment: A small model built for agent work, with tool calling it was trained for, a thinking switch and a 256K context in about 5 GB, under Apache 2.0, from a lab that answers most questions. Its architecture is its own, so check your runtime: SGLang and llama.cpp have it, Transformers needs the repository's code, and Nanbeige ships no GGUF.

## Facts

- Kind: Open-weight model · vendor: Nanbeige LLM Lab (BOSS Zhipin) · category: Open-weight models · legal entity: BOSS Zhipin (Kanzhun Limited), unconfirmed · provenance 58/100
- Packages: huggingface `Nanbeige/Nanbeige4.2-3B`, huggingface `Nanbeige/Nanbeige4.2-3B-FP8`, huggingface `Nanbeige/Nanbeige4.2-3B-GPTQ-Int8`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0 (Hub metadata only; the repository has no LICENSE file)
- Benchmark probes: not run yet, so Performance is pending
- Surface graded: The open weights at huggingface.co/Nanbeige/Nanbeige4.2-3B and the FP8, GPTQ-Int8 and DSpark repositories beside it, read from the model card, the files and their metadata, the config, the tokenizer config with its chat template, the commit history and the runtimes' own code at their latest releases on 10 November 2026. Nanbeige's company facts come from our Nanbeige4.3-33B listing. The technical report in the repository, the repository's Python files and Nanbeige's forks weren't read. Nothing was downloaded or run
- Model: 4,169,800,704 parameters in the files; the card says 3B outside the embeddings. NanbeigeForCausalLM, a looped transformer that reuses its layers, built on Nanbeige4.2-3B-Base with supervised fine-tuning on agent trajectories and reinforcement learning
- Files: Two safetensors shards in BF16 (8.34 GB), config and tokenizer files with the chat template in tokenizer_config.json, two Python modules (configuration_nanbeige.py, modeling_nanbeige.py), the technical report as a PDF and evaluation result files. No pickle files. Hugging Face lists a SHA-256 for the large files
- Licence: Apache 2.0 in the metadata of every repository. There is no LICENSE file and the card's text names no licence
- Runtimes: The card documents Transformers 4.45.1 with trust_remote_code, and SGLang, vLLM, llama.cpp and Ollama from Nanbeige's own forks. Released SGLang 0.5.23 has the model and its tool and reasoning parsers, llama.cpp b12242 has the architecture, and vLLM 0.33.0 names it in its registry and runs it through its Transformers backend. Transformers 5.22.2 has no native model
- Context: 262,144 tokens, per the card and config.json
- Tools and reasoning: Passing tools switches on the tool template, with tool_call_format xml recommended and json supported. enable_thinking is on by default and off per request, and preserve_thinking keeps earlier reasoning for multi-turn tool use. The card's serve commands name the nanbeige tool-call and reasoning parsers
- Hardware: No memory figure on the card. 8.34 GB in BF16 and 5.19 GB in FP8 or 8-bit. The card presents it as a local personal assistant
- Training data: A general description on the card: supervised fine-tuning across real and synthesised tool-use environments with filtering by test cases and rubrics, then reinforcement learning with outcome and process rewards. No sources, sizes or cut-off. The technical report wasn't read

- Scores: Capability pending, Reach 85, Risk 55, Agent interface 85, Documentation and upkeep 72 · total over the 7 assessed categories
- Why:  · Reach, Memory 30, the smallest build from Nanbeige is the FP8 repository (the 8-bit GPTQ one is the same size), 5.19 GB, and 5.19 GB × 1.2 = 6.2 GB… · Risk, File safety 15, safetensors only, no .bin, .pt or .pth files (25). · Agent interface, Chat template 20, A Jinja chat template in tokenizer_config.json (7,961 characters), with enable_thinking, preserve_thinking, tools and tool… · Documentation and upkeep, Model card 17, Intended use, limitations, the architecture, the parameters (3B outside the embeddings), the context and the languages (in th…
- Sources: 14, open questions: 5, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/nanbeige-4-2-3b.json
- Its vendor verifies the listing with `verify_listing` or `POST https://www.anchorterminal.com/api/v1/verify` (the full twin says how); no effect on the grade.
- Support: https://huggingface.co/Nanbeige/Nanbeige4.2-3B/discussions; more in the full twin

## Before you call it

1. Serve it with SGLang 0.5.23 or later and pass `--tool-call-parser nanbeige --reasoning-parser nanbeige`
2. Pass `tools` in the request. The card recommends `tool_call_format="xml"`
3. Set `preserve_thinking` true for multi-turn tool use and coding, false for plain chat
4. Use temperature 1.0 for agent and tool work and 0.6 for reasoning and chat, per the card
5. For llama.cpp use b12242 or later. The GGUF files on the Hub are bartowski's, not Nanbeige's
6. If you load it with Transformers, read modeling_nanbeige.py first: trust_remote_code runs it

## Connect

```bash
pip install "sglang>=0.5.23"
python -m sglang.launch_server --model-path Nanbeige/Nanbeige4.2-3B --reasoning-parser nanbeige --tool-call-parser nanbeige
```

```bash
# SGLang serves an OpenAI-compatible API at http://localhost:30000/v1. Pass tools in the request, and
# chat_template_kwargs {"enable_thinking": false} for a direct answer
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 75.4 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 73.6 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 12B IT | BB | 72.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)