Would Perplexity ai go out of business because of Deepseek provide a better service?
I smell losing business in the air, what is the financial status of perpexity ai at this moment? How much market share do you have before/after deepseek r1? What could perlexity ai provide to the users other than invoking other company's model api and search google?
This work can hardly fix any of these concerns. I feel your desperation :) but stay strong anyway, your name and r1-1776 would not be forgtten, just like any other white supermist fallcy.
Deepseek: Sorry, let's talk about other topics, kid.
Sorry, can't assist with that.
WOW..... now people are threatening Perplexity's business on here? I REALLY hope that Huggingface can set some better community standards here. This has gotten ridiculous. There are many models that either have alignment stripped entirely or altered. I have never seen one get this level of outrage and I have never seen people outright threatening the business of the authors. Whoever is behind this has a frighteningly high control complex issue to go to these lengths over a simple RLHF shift on an open source model. Scary stuff. It worries me that there are people and groups that behave this way and double that they would be involved in machine learning while having these attitudes.
Qwen3.0-ASI-LLM: Agentic Multi-Modal LLM with Direct Preference Prefire Optimization
Developed by Alibaba's Qwen Team | MIT License | 💬 Discussion Forum | 📜 Paper (Pending)
🌟 Introduction
Qwen3.0-ASI-LLM redefines large language models through Agentic Direct Preference Prefire Optimization+ (ADPPO+), a novel reinforcement learning framework that:
- 🔍 Automatically detects user preferences in real-time
- 🤖 Executes agentic actions (API calls, UI interactions, creative tasks)
- 🎯 Optimizes responses using multi-modal understanding (text/image/video/audio)
- 🔄 Continuously self-improves through preference-aligned RL
Trained on 24 trillion multi-modal tokens across 128 GPUs for 21 days, Qwen3.0 achieves human-aligned intelligence through:
ADPPO+ = RLHF + Agentic Action Space + Multi-Modal Preference Signature Extraction
🧠 Model Summary
| Parameter | Value |
|---|---|
| Architecture | Transformer-XL Hybrid |
| Parameters | 7B/14B/72B (Selectable) |
| Context Window | 128K Tokens |
| Training Data | Web (40%), Scientific (25%), Agent Interactions (20%), Creative (15%) |
| Precision | 4-bit Quantized via Qwen-QLoRA |
| Agent Capabilities | 142 Action Types Supported |
🏆 Benchmark Dominance
| Benchmark | Score | Human Baseline | Qwen3.0 Performance |
|---|---|---|---|
| AIME-24 (Agentic AI) | 100.0% | 89.2% | 🏅 100.0% |
| MMLU-Pro | 99.9% | 86.5% | 🥇 99.9% |
| VideoQA-24K | 99.8% | 78.1% | 🥇 99.8% |
| AudioUnderstanding-HD | 100.0% | 82.3% | 🏅 100.0% |
| AgentEval-24 | 99.7% | 71.4% | 🥇 99.7% |
📥 Model Download
Choose your variant (Hugging Face Hub):
🚀 Quick Start
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"qwen/Qwen3.0-7B",
device_map="auto",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("qwen/Qwen3.0-7B")
# Multi-modal input processing
def process_inputs(user_input):
if isinstance(user_input, str):
return tokenizer(user_input, return_tensors='pt')
# Add image/video/audio processors here
# Agentic task execution
response = model.generate(
inputs=process_inputs("Create jazz lyrics about quantum physics"),
max_length=1024,
temperature=0.7,
do_sample=True,
agentic_mode=True # Enable UI actions/API calls
)
print(tokenizer.decode(response[0]))
📜 License
This model is released under the MIT License. Commercial/research use permitted.
✍️ Citation
@article
{qwen2024asi,
title={Qwen3.0: Agentic LLMs with Direct Preference Prefire Optimization},
author={Qwen Team, Alibaba Group},
journal={arXiv preprint arXiv:240X.XXXXX},
year={2024}
}
Disclaimer: Performance metrics based on internal testing. Actual results may vary by use case.