Nanbeige 4.2 (3B, Q4)
Nanbeige LLM Lab's compact agentic reasoning model. A looped transformer — 22 layers run twice over shared weights — so a 3B-parameter download does the compute of a much deeper stack. Text-only, with native thinking, XML tool calling, English + Chinese, and a 256K context window. Q4 builds suit machines with 8GB+ usable memory.
At a glance
- Parameters
- 3B
- Quantization
- Q4_K_M
- Context window
- 262K tokens
- Approx. size
- 2.7 GB
- Engines
- llama.cpp
- License
- Apache-2.0 (open)
- Version
- v1.0.0 , released 2026-08-31
- Family
- other
Tuning
Defaults a gezel client applies out of the box, from real evaluation runs against this exact quantization.
- sampling.temperature
0.6- sampling.topP
0.95- sampling.topK
20- sampling.maxTokens
16384- reasoning.thinkingBudget
8192- reasoning.enableThinking
true- profiles
- thinking-general, thinking-coding, thinking-precise, instruct, creative
Model behaviors
Client-side behavior modules the catalog enables for this model (reasoning-tag handling, fabrication detection, prompt shaping).
- provider.merge-system-messages
- reasoning.strip-think-tags
- prompt.private-reasoning-guidance
- prompt.tool-cookbook-full
- mcp.compact-tool-schemas
- mcp.relax-required-fields
- mcp.default-missing-fields
- fabrication.detect-past-tense-no-tools
- prompt.retrieval-first
- turn.continuation-budget
- turn.preamble-folding
- turn.ramble-detection
- prompt.prefer-writefile-edits
Sources
- llama.cpp (GGUF)
- bartowski/Nanbeige_Nanbeige4.2-3B-GGUF ·
Nanbeige_Nanbeige4.2-3B-Q4_K_M.gguf· Q4_K_M - Upstream
- https://huggingface.co/Nanbeige/Nanbeige4.2-3B
- Manifest
- View on GitHub
Tags
- nanbeige
- reasoning
- agentic
- coding
- tools
- long-context
- multilingual
- small
- open-weights
- apple-silicon