What is Nanbeige4.2-3B?
Nanbeige4.2-3B is a small agent-focused model from the Nanbeige LLM Lab at BOSS Zhipin, the Chinese recruitment platform, released in July 2026. It has 3 billion non-embedding parameters (4 billion in total) and uses a looped transformer: the same layers run more than once per token, adding capacity without adding weights. It reads and writes text, in English and Chinese, with a 256K-token context.
It was fine-tuned on a wide mix of real and synthetic agent environments, then trained with reinforcement learning on both final outcomes and intermediate steps. Thinking is on by default and can be switched off. The lab's larger Nanbeige4.3-33B followed in October; the 3B remains its current small model, and Nanbeige says a 4.5 release is planned for later in 2026.
The weights are under Apache 2.0. In the last 30 days they drew about 121,000 Hugging Face downloads.
Is Nanbeige4.2-3B worth it?
Nanbeige4.2-3B posts agent scores that belong to a model three times its size. On the lab's own table it scores 63.6% on SWE-Bench Verified against 53.1% for Qwen3.5-9B and 44.2% for Gemma 4 12B, 44.1 on Terminal-Bench 2.0 against 29.2, and 57.8 on MCP-Atlas tool use against 47.4. Reasoning holds up too: 87.4 on GPQA Diamond and 82.8 on HMMT February 2026, both ahead of Qwen3.5-9B.
One independent result backs the general picture. Artificial Analysis's phone test in August 2026 put it joint first among small models at 63 with LFM2.5-2.6B, with 76% on BFCL function calling and 96% on MATH-500. The same test found the weak spot: only 33% on its non-hallucination measure against 79% for LFM2.5-2.6B, so it makes up facts more readily than rivals.
The bigger catch is how long it thinks. All of Nanbeige's scores are in thinking mode, with up to 131,072 new tokens recommended for reasoning. When Artificial Analysis capped each answer at one minute on an iPhone 17 Pro, its score fell from 63 to 18, because at about 14 tokens a second most of its reasoning never finished. A 1,024-token prompt with a 256-token reply took 21.4 seconds in 4.0 GB, against 8.0 seconds and 2.3 GB for LFM2.5-2.6B.
So it suits laptops and desktop GPUs running agents in the background, where waiting is fine. It also needs trust_remote_code and Nanbeige's own SGLang fork, which adds review work for security-minded teams. Gemma 4 12B still leads it on IF-Bench instruction following, 73.5 to 54.6.
Where it earns its place
A solo developer maintaining several open-source libraries
Have a local agent attempt small bug fixes overnight without paying for API tokens
Serve it with Nanbeige's SGLang fork and point OpenHands at it with preserve_thinking on. Its 63.6% SWE-Bench Verified score is self-reported; review every patch before merging.
An office manager at a 20-person accounting practice with an IT contractor
Run a private assistant for document lookup and routine office tasks
Run it on a desktop GPU through an agent framework such as OpenClaw, where Nanbeige reports 68.8 on GDPval against 38.0 for Qwen3.5-9B. Its 33% non-hallucination score means checking any figures it quotes.
Before you commit
- Setup
- Install transformers 4.45.1 and load with trust_remote_code, or clone Nanbeige's SGLang branch for serving. Set enable_thinking and preserve_thinking in the chat template; XML tool-call format is recommended.
- Integrations
- transformers, Nanbeige's SGLang fork, OpenHands, SWE-agent and Terminus scaffolds in its tests, OpenClaw for personal workflows, and a DSpark draft model for speculative decoding.
- Data and privacy
- Self-hosted, so prompts and files stay on your hardware. Apache 2.0 permits commercial use. Custom modelling code runs with trust_remote_code, so review it before deployment.
- Cost at scale
- Weights are free. Artificial Analysis measured 4.0 GB of memory on an iPhone 17 Pro; long thinking traces mean far more tokens per task than a non-reasoning model.
- Limits
- About 14 tokens a second on a phone; recommended reasoning budgets run to 131,072 tokens; English and Chinese only.
What’s good
- 63.6% on SWE-Bench Verified, above Qwen3.5-9B (self-reported)
- Joint first in Artificial Analysis's small-model phone test
- 256K context and tool calling at 3B non-embedding parameters
- Apache 2.0
What’s not
- Long thinking: scored 18 under a one-minute cap
- 33% non-hallucination in Artificial Analysis's test
- Needs trust_remote_code and a forked SGLang
- Trails Gemma 4 12B on IF-Bench instruction following
Best for
Developers running a local coding or office agent on a laptop or single GPU who can tolerate long reasoning times.
Nanbeige4.2-3B alternatives
Nanbeige4.2-3B: common questions
- Is Nanbeige4.2-3B worth it?
- Nanbeige4.2-3B does agent work that used to need a 9B model, provided you give it time to think. We rate it 4.2 out of 5.
- How much does Nanbeige4.2-3B cost?
- Free Apache 2.0 weights. No first-party hosted API published.
- What are the downsides of Nanbeige4.2-3B?
- Long thinking: scored 18 under a one-minute cap; 33% non-hallucination in Artificial Analysis's test; needs trust_remote_code and a forked SGLang; trails Gemma 4 12B on IF-Bench instruction following.
- Who is Nanbeige4.2-3B best for?
- Developers running a local coding or office agent on a laptop or single GPU who can tolerate long reasoning times.
- How much does Nanbeige4.2-3B cost for a team or in production?
- Weights are free. Artificial Analysis measured 4.0 GB of memory on an iPhone 17 Pro; long thinking traces mean far more tokens per task than a non-reasoning model.
- Is Nanbeige4.2-3B safe for company data?
- Self-hosted, so prompts and files stay on your hardware. Apache 2.0 permits commercial use. Custom modelling code runs with trust_remote_code, so review it before deployment.
- What are the best Nanbeige4.2-3B alternatives?
- The closest alternatives we have reviewed are Spark-X2.6-4B, LFM3, MiniCPM5-2B, Gemma 4 and Nanbeige4.3-33B.
Answers checked Nov 7, 2026.
Make Nanbeige4.2-3B? Show this rating on your site
