GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
nabin2004
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
nabin2004
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
| Role | Artifact / Repository | Description |
|---|---|---|
| Base LLM | Qwen/Qwen3-8B | Base causal foundation model |
| SFT Prior | nabin2004/AOS-qwen3-8b-narrated-adapter | Continued SFT on 400 synchronized educational voiceover trajectories |
| DPO Adapter | nabin2004/AOS-qwen3-8b-narrated-dpo | Direct Preference Optimization adapter (β=0.1) |
| Merged Weights | nabin2004/AOS-qwen3-8b-narrated-merged | Full unquantized bfloat16 Safetensors weights (this repo) |
| GGUF / Ollama | nabin2004/AOS-qwen3-8b-narrated-gguf | Multi-quantized GGUF (Q4_K_M, Q8_0) for Ollama & llama.cpp |
The model was aligned with Direct Preference Optimization (DPO) to strongly prefer generating voiceover-synchronized educational animations:
VoiceoverScene scripts with speech services (AOSSpeechService / GTTSService), animation duration tracking (run_time=tracker.duration), millisecond-accurate <bookmark mark='...'/> tags, and natural phonetic spoken narration.Scene code.python
from manim import *from manim_voiceover import VoiceoverScenefrom manim_voiceover.services.gtts import GTTSServiceclass SigmoidExplanation(VoiceoverScene):def construct(self):# Configure speech serviceself.set_speech_service(GTTSService())title = Title("The Sigmoid Activation Function")ax = Axes(x_range=[-6, 6, 2], y_range=[-0.2, 1.2, 0.5])curve = ax.plot(lambda x: 1 / (1 + np.exp(-x)), color=BLUE)dot = Dot(ax.c2p(0, 0.5), color=RED)with self.voiceover(text="Let's visualize the sigmoid function. <bookmark mark='AXES'/> We begin by setting up our coordinate system, <bookmark mark='CURVE'/> plotting the characteristic S-shaped curve, <bookmark mark='DOT'/> and marking the midpoint inflection at zero, point five.") as tracker:self.play(Write(title))self.wait_until_bookmark("AXES")self.play(Create(ax))self.wait_until_bookmark("CURVE")self.play(Create(curve))self.wait_until_bookmark("DOT")self.play(FadeIn(dot), run_time=tracker.duration)self.wait(1)
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizermodel_id = "nabin2004/AOS-qwen3-8b-narrated-merged"tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained(model_id,torch_dtype=torch.bfloat16,device_map="auto",trust_remote_code=True,)prompt = "Create a narrated Manim animation explaining the Fourier Transform with voiceover bookmarks."messages = [{{"role": "system", "content": "You are an expert mathematical animation assistant specializing in Manim Community Edition and voiceover narration with manim-voiceover. You write complete, self-contained, fully executable Python scripts inheriting from VoiceoverScene."}},{{"role": "user", "content": prompt}}]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=2048,temperature=0.2,top_p=0.95,pad_token_id=tokenizer.eos_token_id,)print(tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True))
Serve as a high-performance OpenAI-compatible endpoint:
bash
vllm serve nabin2004/AOS-qwen3-8b-narrated-merged \--port 8000 \--max-model-len 8192 \--trust-remote-code
Part of the AOS (Agentic Orchestration System) project for multi-agent educational video synthesis.
Qwen/Qwen3-8B)manim-voiceover