A tiny (~42 MB) randomly-initialized version of MiniCPM-o-2.6 designed for testing purposes in the optimum-intel library.
Purpose
This model was created to replace the existing test model at optimum-intel-internal-testing/tiny-random-MiniCPM-o-2_6 (185 MB) with a smaller alternative for CI/CD testing. Smaller test models reduce:
Result: 4× smaller than Intel's current test model
Model Configuration
Component
This Model
Original
Vocabulary
5,000 tokens
151,700 tokens
LLM Hidden Size
128
3,584
LLM Layers
1
40
LLM Attention Heads
8
28
Vision Hidden Size
128
1,152
Vision Layers
1
27
Image Size
980 (preserved)
980
Patch Size
14 (preserved)
14
Audio d_model
64
1,280
TTS Hidden Size
128
-
Parameter Breakdown
Component
Parameters
Size (MB)
TTS/DVAE
19,339,766
36.89
LLM
1,419,840
2.71
Vision
835,328
1.59
Resampler
91,392
0.17
Audio
56,192
0.11
Other
20,736
0.04
Total
21,763,254
~41.5
Technical Details
Why Keep TTS/DVAE Components?
The TTS (Text-to-Speech) component, which includes the DVAE (Discrete Variational Auto-Encoder), accounts for approximately 37 MB (~85%) of the model size. While the optimum-intel tests do not exercise TTS functionality (they only test image+text → text generation), we retain this component because:
Structural Consistency: Removing TTS via init_tts=False causes structural differences in the model that lead to numerical divergence between PyTorch and OpenVINO outputs
Test Compatibility: The test_compare_to_transformers test compares PyTorch vs OpenVINO outputs and requires exact structural matching
Architecture Integrity: The MiniCPM-o architecture expects TTS weights to be present during model loading
Tokenizer Shrinking
The vocabulary was reduced from 151,700 to 5,000 tokens:
Base tokens: IDs 0-4899 (first 4,900 most common tokens)
Special tokens: IDs 4900-4949 (remapped from original high IDs)
BPE merges: Filtered from 151,387 to 4,644 (only merges involving retained tokens)
Key token mappings:
Token
ID
<unk>
4900
<|endoftext|>
4901
<|im_start|>
4902
<|im_end|>
4903
Reproducibility
Model weights are initialized with a fixed random seed (42) to ensure:
Reproducible outputs between runs
Consistent behavior between PyTorch and OpenVINO
Passing of test_compare_to_transformers which compares framework outputs
Test Results
Tested with pytest tests/openvino/test_seq2seq.py -k "minicpmo" -v:
Test
Status
Notes
test_compare_to_transformers
✅ PASSED
PyTorch/OpenVINO outputs match
test_generate_utils
✅ PASSED
Generation pipeline works
test_model_can_be_loaded_after_saving
⚠️ FAILED
Windows file locking issue (not model-related)
The third test failure is a Windows-specific issue where OpenVINO keeps file handles open, preventing cleanup of temporary directories. This is a known platform limitation, not a model defect. The test passes on Linux/macOS.
Usage
For optimum-intel Testing
python
1# In optimum-intel/tests/openvino/utils_tests.py, update MODEL_NAMES:2MODEL_NAMES ={3# ... other models ...4"minicpmo":"hrithik-dev8/tiny-random-MiniCPM-o-2_6",5}
