A supervised fine-tune of openbmb/MiniCPM5-2B on sriq-ai/sriq-sft-v1.5, by SRIQ.
The model reasons in compressed Simplified Chinese and answers in the language it was asked in. The reasoning is telegraphic — no hedging, no restating the question, no thinking out loud — and closes by checking the result against the constraints it started from. The final answer is unaffected: full length, in the reader's language.
Chinese carries a reasoning step in roughly half the tokens English needs, so the trace costs less without the answer getting shorter. That is the point of the model.
Output shape
Reasoning is emitted inside <think> tags, then the answer:
<think>设 s=x+y，p=xy。…验：23*17=391。答案对。</think>
x + y = 2 or 5. Working: …
v1.5's training data ends every trace with a verification block — substituting the result back, re-deriving the invariant, walking the stated edge cases — before the answer is committed.
Reasoning under an agent harness
v1.5 is the dataset version that adds long system prompts simulating agentic coding harnesses. Earlier versions held Chinese reasoning under a short system prompt but drifted back to English behind a real agent harness — a multi-thousand-token English preamble of tool descriptions, skill listings and project context. 599 of the dataset's 1,424 rows carry such a preamble, so the behaviour is trained in the context where it otherwise collapses.
Usage
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "sriq-ai/sriq-MiniCPM5-2B-v1.5"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")
messages = [{"role": "user", "content": "What is 17 * 23? Answer briefly."}]
inputs = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
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))
Do not pass enable_thinking=False
The chat template treats that flag as an instruction to skip reasoning, and prefills an empty block:
{%- if enable_thinking is false %}
{{- '<think>\n\n</think>\n\n' }}
The model then answers with no trace at all, silently — there is no error, and the behaviour this
model exists for is switched off. Leave the flag undefined, or set it to True.
Model details
| Base model | openbmb/MiniCPM5-2B |
| Parameters | 2,516,756,480 |
| Architecture | LlamaForCausalLM |
| Weights | bfloat16, merged (not an adapter) |
| Context length | 131,072 |
| Training data | sriq-ai/sriq-sft-v1.5 — 1,424 conversations, 30 languages, all difficulty: hard |
| Method | Supervised fine-tuning with Unsloth |
Quantizations: sriq-MiniCPM5-2B-v1.5-GGUF.
Evaluation
No benchmark results are published for this model. Nothing here describes measured gains in accuracy, verification rate, or answer quality — only what it was trained on and how it behaves. Evaluate on your own workload before relying on it.
Limitations
- The reasoning is Chinese regardless of the question's language. If you show traces to users, they will see Chinese.
- Compressed traces are harder to read than narrated ones. That is the trade being made.
- A 2B model is a 2B model: the verification step checks the working, it does not make the working correct.
License
Apache-2.0, inherited from the base model.