GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ImposterOnline
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
ImposterOnline
Model Tree
BlossomsAI/Qwen2.5-Coder-7B-Instruct-Uncensored
Input Modalities
Output Modalities
Supported Functionality
Pyrex 8B Instruct Uncensored is a real, trained model — built with our own data recipe, our own QLoRA fine-tune, and our own evaluation pipeline on Hugging Face GPU compute. It is a coding-first instruction model designed for real developer work: writing code, fixing bugs, explaining technical problems, and driving agent loops.
Uncensored usage note. This model has no safety-alignment softening built in. It can produce content that is explicit, offensive, or otherwise unsuitable for some audiences. Use it responsibly and at your own discretion — the model and its authors assume no liability for outputs.
| Property | Value |
|---|---|
| Parameters | 7.6B |
| Architecture | Transformers · RoPE · SwiGLU · RMSNorm · GQA |
| Context window | Up to 32k native (trained at 3072) |
| Base | Coding-specialized instruct model (uncensored variant) |
| Method | QLoRA (4-bit NF4) + LoRA r=48, α=96 |
| Chat format | chatml (`< |
| License | Apache-2.0 |
Measured with greedy pass@1 on HumanEval — 164 unseen problems, official unit tests executed in a sandbox. No sampling luck, no leaked answers.
| Model | HumanEval pass@1 |
|---|---|
| Base (untuned) | 29.3% (48/164) |
| Pyrex (previous) | 41.5% (68/164) |
| Pyrex 8B Instruct Uncensored | 52.4% (86/164) |
+23.1 points over the base on unseen problems.
transformers >= 4.44 for best apply_chat_template support.python
from transformers import AutoModelForCausalLM, AutoTokenizermodel_id = "ImposterOnline/Pyrex-8B-Instruct-Uncensored"tok = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype="auto")messages = [{"role": "user", "content": "Explain async/await in Python."}]text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)out = model.generate(**tok(text, return_tensors="pt"), max_new_tokens=512)print(tok.decode(out[0], skip_special_tokens=True))
Chat format is chatml. System prompt is optional; when used it is a system
message before the user turn.
Works with any OpenAI-compatible server (vLLM, llama.cpp server, Ollama, Text Generation Inference). Function calling was part of training.
bash
vllm serve ImposterOnline/Pyrex-8B-Instruct-Uncensored
Quantized builds live in Pyrex-8B-Instruct-Uncensored-GGUF.
| File | Size | Notes |
|---|---|---|
pyrex-q4_k_m.gguf | 4.4 GB | Fast, recommended daily driver |
pyrex-q5_k_m.gguf | 5.1 GB | Higher quality |
pyrex-q8_0.gguf | 8.1 GB | Fast, best quality |
pyrex-f16.gguf | 15.2 GB | Full precision |
bash
# Ollama (Modelfile ships in the GGUF repo)ollama create pyrex-8b -f Modelfileollama run pyrex-8b "Write a Python function that merges overlapping intervals."# llama.cpp./llama-cli -m pyrex-q4_k_m.gguf \-p "<|im_start|>user\nYour question here\n<|im_end|>\n<|im_start|>assistant\n" \-n 512 -c 8192
| Parameter | Value |
|---|---|
| Data | ~95k cleaned examples (code + tool-use + general) |
| Sequence length | 3072 |
| Loss masking | Assistant-only (completion-only SFT) |
| Optimizer | AdamW · lr 2e-4 · cosine · warmup 3% |
| Epochs | 1 |
| Hardware | Hugging Face A100 GPU job |
Data sources (credited): OpenCoder-LLM/opencoder-sft-stage1
(realuser + largescale-diverse), sahil2801/CodeAlpaca-20k,
theblackcat102/evol-codealpaca-v1, databricks/databricks-dolly-15k,
NousResearch/hermes-function-calling-v1.
The pipeline is config-driven and reproducible — prepare_data.py →
train_qlora.py → eval_humaneval.py → publish.sh — in the companion
repo Arhan-w/pyrut.
Apache-2.0.