MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic-heretic-ara-strict API & Inference Endpoint

Abliteration parameters

Table with columns: Parameter, Value
Parameter	Value
start_layer_index	0
end_layer_index	33
preserve_good_behavior_weight	0.7812
steer_bad_behavior_weight	0.0001
overcorrect_relative_weight	0.0848
neighbor_count	8

Performance

Table with columns: Metric, This model, Original model (GnLOLot/MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic)
Metric	This model	Original model (GnLOLot/MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic)
KL divergence	0.0052	0 (by definition)
Refusals	2/100	90/100

MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic

GGUF quantizations for local deployment: MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic-GGUF

MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic is a compact 2B Thinking language model built on openbmb/MiniCPM5-2B. Fine-tuned on Claude data with a strong focus on agentic tool calling / function calling, coding, and instruction following. It keeps MiniCPM5's native Thinking chat template and XML tool-call format.

For llama.cpp / Ollama / LM Studio deployment, see the GGUF repository.

Overview

Table with columns: Item, Detail
Item	Detail
Base model	openbmb/MiniCPM5-2B (2B dense Llama architecture)
Post-training	Claude data
Key capabilities	Agentic tool calling, coding, instruction following, chain-of-thought reasoning
Chat format	MiniCPM5 native Thinking template with optional chain-of-thought blocks
Context length	128K (`max_position_embeddings = 131072`)
Precision	bfloat16

Capabilities

Agentic tool calling — reliable XML / function-calling style tool use on top of MiniCPM5's native format, designed for multi-step agentic workflows
Coding — code generation, debugging, and software-engineering-style tasks
Instruction following — reliable adherence to user prompts and structured constraints
Thinking mode — chain-of-thought reasoning via the MiniCPM5 chat template
Long context — up to 128K tokens (131,072 tokens per config.json)

Benchmark

ClawBench (Agentic Coding)

Table with columns: Model, QwenClawBench, WildClawBench
Model	QwenClawBench	WildClawBench
MiniCPM5-2B (Base, RL-only)	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. QwenClawBench uses structured coding scenarios; WildClawBench tests on diverse real-world tasks.

More benchmarks (BFCL, SWE-bench, Tau-Bench, etc.) coming soon.

Quick start

python

from transformers import AutoModelForCausalLM, AutoTokenizerimport torch model_id = "GnLOLot/MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic" tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained(    model_id,    trust_remote_code=True,    torch_dtype=torch.bfloat16,    device_map="auto",) messages = [{"role": "user", "content": "Write a Python function to merge two sorted lists."}]text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(text, return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Tool calling example

python

tools = [    {        "type": "function",        "function": {            "name": "get_weather",            "description": "Get the current weather for a given city.",            "parameters": {                "type": "object",                "properties": {                    "city": {"type": "string", "description": "City name"}                },                "required": ["city"]            }        }    }] messages = [    {"role": "user", "content": "What's the weather like in Beijing?"}] text = tokenizer.apply_chat_template(    messages, tools=tools, tokenize=False, add_generation_prompt=True)inputs = tokenizer(text, return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Sampling recommendations

Inherited from openbmb/MiniCPM5-2B:

Table with columns: Scenario, Params
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.

Limitations

Thinking outputs — the model may emit reasoning blocks before the final answer; downstream apps can strip them before display
2B scale — optimized for lightweight local deployment, not frontier-scale general reasoning

Provenance & licensing

Released under Apache-2.0, inherited from MiniCPM5-2B.

Acknowledgements

Base model: OpenBMB / MiniCPM5-2B
GGUF conversion: llama.cpp

Abliteration parameters

Table with columns: Parameter, Value
Parameter	Value
start_layer_index	0
end_layer_index	33
preserve_good_behavior_weight	0.7812
steer_bad_behavior_weight	0.0001
overcorrect_relative_weight	0.0848
neighbor_count	8

Performance

Table with columns: Metric, This model, Original model (GnLOLot/MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic)
Metric	This model	Original model (GnLOLot/MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic)
KL divergence	0.0052	0 (by definition)
Refusals	2/100	90/100

MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic

GGUF quantizations for local deployment: MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic-GGUF

For llama.cpp / Ollama / LM Studio deployment, see the GGUF repository.

Overview

Table with columns: Item, Detail
Item	Detail
Base model	openbmb/MiniCPM5-2B (2B dense Llama architecture)
Post-training	Claude data
Key capabilities	Agentic tool calling, coding, instruction following, chain-of-thought reasoning
Chat format	MiniCPM5 native Thinking template with optional chain-of-thought blocks
Context length	128K (`max_position_embeddings = 131072`)
Precision	bfloat16

Capabilities

Agentic tool calling — reliable XML / function-calling style tool use on top of MiniCPM5's native format, designed for multi-step agentic workflows
Coding — code generation, debugging, and software-engineering-style tasks
Instruction following — reliable adherence to user prompts and structured constraints
Thinking mode — chain-of-thought reasoning via the MiniCPM5 chat template
Long context — up to 128K tokens (131,072 tokens per config.json)

Benchmark

ClawBench (Agentic Coding)

Table with columns: Model, QwenClawBench, WildClawBench
Model	QwenClawBench	WildClawBench
MiniCPM5-2B (Base, RL-only)	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. QwenClawBench uses structured coding scenarios; WildClawBench tests on diverse real-world tasks.

More benchmarks (BFCL, SWE-bench, Tau-Bench, etc.) coming soon.

Quick start

python

from transformers import AutoModelForCausalLM, AutoTokenizerimport torch model_id = "GnLOLot/MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic" tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained(    model_id,    trust_remote_code=True,    torch_dtype=torch.bfloat16,    device_map="auto",) messages = [{"role": "user", "content": "Write a Python function to merge two sorted lists."}]text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(text, return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Tool calling example

python

tools = [    {        "type": "function",        "function": {            "name": "get_weather",            "description": "Get the current weather for a given city.",            "parameters": {                "type": "object",                "properties": {                    "city": {"type": "string", "description": "City name"}                },                "required": ["city"]            }        }    }] messages = [    {"role": "user", "content": "What's the weather like in Beijing?"}] text = tokenizer.apply_chat_template(    messages, tools=tools, tokenize=False, add_generation_prompt=True)inputs = tokenizer(text, return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Sampling recommendations

Inherited from openbmb/MiniCPM5-2B:

Table with columns: Scenario, Params
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.

Limitations

Thinking outputs — the model may emit reasoning blocks before the final answer; downstream apps can strip them before display
2B scale — optimized for lightweight local deployment, not frontier-scale general reasoning

Provenance & licensing

Released under Apache-2.0, inherited from MiniCPM5-2B.

Acknowledgements

Base model: OpenBMB / MiniCPM5-2B
GGUF conversion: llama.cpp

MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic-heretic-ara-strict

README

Abliteration parameters

Performance

MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic

Overview

Capabilities

Benchmark

ClawBench (Agentic Coding)

Quick start

Tool calling example

Sampling recommendations

Limitations

Provenance & licensing

Acknowledgements

Explore FriendliAI today

README

Abliteration parameters

Performance

MiniCPM5-2B-Claude-Fable5-1-Thinking-Agentic

Overview

Capabilities

Benchmark

ClawBench (Agentic Coding)

Quick start

Tool calling example

Sampling recommendations

Limitations

Provenance & licensing

Acknowledgements