MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic-GGUF
GGUF quantizations of MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic for llama.cpp, Ollama, LM Studio, jan, KoboldCpp, and other GGUF runtimes.
This repository provides local-deployment builds of a 2B Thinking model fine-tuned on Claude data atop openbmb/MiniCPM5-2B, with a strong focus on agentic tool calling, coding, and instruction following. MiniCPM5's native chat template is embedded in the GGUF files.
Transformers checkpoint: MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic
Files
| File | Quant | Size | Notes |
|---|---|---|---|
MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic-F16.gguf | F16 | ~4.7 GB | full-precision conversion base |
MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic-Q8_0.gguf | Q8_0 | ~2.5 GB | recommended default |
MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic-Q4_K_M.gguf | Q4_K_M | ~1.5 GB | smallest, for memory-constrained devices |
Q8_0 is the recommended default quant for this 2B model. Q4_K_M is available for ultra-lightweight deployment.
Quick start
llama.cpp (llama-cli)
llama-cli \
-m MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic-Q8_0.gguf \
-p "Write a Python function to merge two sorted lists." \
-n 512 \
--temp 1.0 --top-p 0.95 --min-p 0.0 \
-c 8192
The model supports up to 128K tokens (131,072) per
config.json. Set-caccording to your available VRAM/RAM.
llama.cpp server
llama-server \
-m MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic-Q8_0.gguf \
-c 8192 --port 8080
LM Studio / jan / KoboldCpp
Load any .gguf file from this repository. The MiniCPM5 chat template is embedded in the GGUF metadata.
Sampling recommendations
Inherited from openbmb/MiniCPM5-2B:
| Scenario | Params |
|---|---|
| Default | temperature=1.0, top_p=0.95, min_p=0.0 |
| If repetitive outputs | temperature=1.0, top_p=0.95, min_p=0.0, repetition_penalty=1.05 |
This model is Thinking-only — chain-of-thought reasoning is always active.
Support for sampling parameters varies across inference frameworks — check your runtime's documentation.
Capabilities
- Agentic tool calling — reliable function-calling / tool-use behavior designed for multi-step agentic workflows
- Claude fine-tune — post-trained on Claude data
- Coding — code generation, debugging, and software-engineering workflows
- Instruction following — reliable adherence to user prompts and task constraints
- Thinking mode — chain-of-thought reasoning; MiniCPM5 chat template baked into the GGUF
- Long context — up to 128K tokens (131,072 tokens per upstream
config.json)
Benchmark
Scores for the Transformers checkpoint MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic:
ClawBench (Agentic Coding)
| Model | QwenClawBench | WildClawBench |
|---|---|---|
| MiniCPM5-2B | 42.11 | 23.19 |
| MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic | 44.56 (+2.45) | 24.32 (+1.13) |
ClawBench evaluates agentic coding ability — the model's capacity to autonomously use tools, navigate codebases, and complete multi-step software engineering tasks.
More benchmarks (BFCL, SWE-bench, Tau-Bench, etc.) coming soon.
Limitations
- Thinking outputs — the model may emit reasoning blocks before the final answer
- 2B scale — lightweight local deployment; not frontier-scale
- Runtime context — actual usable context depends on your GGUF runtime and hardware limits
Provenance & licensing
Apache-2.0, inherited from MiniCPM5-2B.
Acknowledgements
- Base model: OpenBMB / MiniCPM5-2B
- Transformers checkpoint: MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic
- Quantization: llama.cpp