Open-weights models,
under Apache 2.0
Our models are on Hugging Face, and the licence lets you run, test, change and ship them in your own products.
Matilda Platypus · Oct 2026
Eight sizes between the Qwen3 models.
Each joins the lower layers of a larger Qwen3 base model to the upper layers of a smaller one, so you can pick the size that fits your hardware.
The collection on Hugging Face- Licence
- Apache 2.0
- Built on
- Qwen3 base models8B and 4B for the 7B member
- Type
- Base modelsThey continue text, with no instruction or safety tuning
- Library
transformers4.56 or laterwith remote code
Pick a size
7B on Hugging Facefrom transformers import AutoModelForCausalLM, AutoTokenizer
repo = "Maincode/Matilda-Platypus-7B-Base"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)Smaller modelupper layers
Representation Bridgelearned, linear
Larger modellower layers
Diagram of the larger Qwen3 model’s lower layers at the bottom, the Representation Bridge in the middle, and the smaller Qwen3 model’s upper layers on top.
Table view
| Benchmark | Qwen3-4B-Base | Platypus 7B | Qwen3-8B-Base |
|---|---|---|---|
| HumanEval pass@1 | 0.628 | 0.689 | 0.677 |
| GSM8K (8-shot CoT) | 0.844 | 0.875 | 0.862 |
| MMLU (5-shot) | 0.730 | 0.763 | 0.768 |
| BBH (3-shot CoT) | 0.712 | 0.746 | 0.774 |
- 11 of 13
- benchmarks above the smaller parent
- 3 of 13
- above the larger parent
- 84%
- of the larger parent’s compute
Scored as proportions with the base-model protocol in lm-eval-harness 0.4.13. Compute is the technical report’s relative measure, not a measured latency or cost, and the full tables are on the model card.
Maincoder-1B · Dec 2025
A 1B-parameter model for Python code.
Maincoder-1B generates and completes Python, and it scores above every baseline on its model card, including SmolLM3-3B at three times its size.
- Licence
- Apache 2.0
- Size
- 1B parameters32 layers, 2,048-token context
- Architecture
- Transformer decoderwith RoPE, grouped-query attention, QK normalisation and SwiGLU
- Fine-tuning
- MCPOa specialised reinforcement-learning policy optimisation algorithm
From the model card, against every baseline it reports.
Table view
| Model | HumanEval | HumanEval+ | MBPP+ |
|---|---|---|---|
| Maincoder-1B | 0.762 | 0.726 | 0.709 |
| deepseek-coder-1.3b-instruct | 0.561 | 0.531 | 0.622 |
| Qwen2.5-Coder-1.5B-Instruct | 0.463 | 0.445 | 0.656 |
| SmolLM3-3B | 0.537 | 0.500 | 0.680 |
| Qwen3-1.7B | 0.402 | 0.378 | 0.558 |
from transformers import (
AutoModelForCausalLM,
AutoTokenizer,
)
repo = "Maincode/Maincoder-1B"
tokenizer = AutoTokenizer.from_pretrained(
repo, trust_remote_code=True
)
model = AutoModelForCausalLM.from_pretrained(
repo, trust_remote_code=True
)1B parameters, small enough for local and on-device inference
More from the lab
