Qwen3

Top Builders

Explore the top contributors showcasing the highest number of app submissions within our community.

Qwen3

Qwen3 is the third-generation text model family from Alibaba Cloud's Qwen team, released on April 28, 2025. It covers six dense sizes (0.6B to 32B) and two MoE variants, all trained on approximately 36 trillion tokens across 119 languages. A key design choice is a unified thinking and non-thinking mode in every model, so developers can choose between step-by-step reasoning and fast single-pass responses without switching models.

General
Release date	28 Apr 2025
Developer	Qwen / Alibaba Cloud
Type	Open-weight text LLM family
License	Apache 2.0
GitHub	QwenLM/Qwen3
Hugging Face	huggingface.co/Qwen
Documentation	qwenlm.github.io/blog/qwen3

Core Features

Thinking/non-thinking mode: every model supports both step-by-step chain-of-thought reasoning and direct response generation within a single checkpoint.
Thinking budget: developers can set a token budget for the reasoning phase, allowing inference cost to be tuned per request.
Long context: models at 4B and above support 131,072-token context windows; 0.6B and 1.7B support 32,768 tokens.
Multilingual: pretrained on 119 languages and dialects.
Apache 2.0: all weights are released for commercial use, fine-tuning, and redistribution.

Model Variants

Variant	Total Params	Active Params	Context	Best for
Qwen3-0.6B	0.6B	0.6B	32K	Edge and on-device
Qwen3-1.7B	1.7B	1.7B	32K	Lightweight inference
Qwen3-4B	4B	4B	128K	Balanced performance
Qwen3-8B	8B	8B	128K	General tasks
Qwen3-14B	14B	14B	128K	Higher accuracy
Qwen3-32B	32B	32B	128K	Strong reasoning
Qwen3-30B-A3B	30B	3B	128K	Efficient MoE
Qwen3-235B-A22B	235B	22B	128K	Flagship MoE

Benchmarks

The flagship Qwen3-235B-A22B model scores:

AIME '24: 85.7
AIME '25: 81.5
LiveCodeBench v5: 70.7
BFCL v3: 70.8

Tools and Resources

Qwen3 Blog Post: official release notes and technical details.
Hugging Face (Qwen): download model weights and GGUF files.
Technical Report: arXiv paper covering architecture, training, and evaluation.
Ollama: run Qwen3 models locally.
Qwen API Platform: access models via OpenAI-compatible API.

Ecosystem and Integrations

Available on Hugging Face Hub in both standard and GGUF formats.
Accessible via Alibaba Cloud DashScope using an OpenAI-compatible endpoint.
Supported by Ollama, LM Studio, and major inference frameworks including vLLM and llama.cpp.
All sizes available for fine-tuning using standard supervised fine-tuning and RL pipelines.

Qwen3 weights are available immediately on Hugging Face. To access via API, generate a key on the Qwen API Platform and follow the Model Studio documentation.

Edit on GitHub

Qwen Qwen3 AI technology Hackathon projects

Discover innovative solutions crafted with Qwen Qwen3 AI technology, developed by our community members during our engaging hackathons.

Project Hermes

Hermes is an Android accessibility and translation assistant designed for real-time communication across text, camera, and voice workflows. The app is built in Kotlin with Android native UI components, CameraX for live camera preview, ML Kit Text Recognition for on-device OCR, Android SpeechRecognizer for fast speech-to-text, and ExecuTorch/QNN integration for experimenting with local AI inference using whisper and Qwen2.5-0.5B and Qwen3-7B and on Qualcomm-powered devices. Its camera mode lets a user point the phone at signs, notes, labels, forms, menus, or other printed text, recognize the text on-device, identify the source language, and translate it into English. Its voice mode supports quick speech transcription so spoken input can be converted into readable text during live interaction. The project is especially useful in real-world situations such as travel, hospitals, classrooms, public offices, accessibility support, and multilingual conversations where users need quick understanding without switching between multiple apps. Hermes also includes groundwork for local model execution using Whisper-style ASR assets, Qwen translation models, Android native libraries, and Qualcomm acceleration paths, making it a practical prototype for privacy-aware, low-latency mobile AI translation.

SixthSense: Haptic Vision for the Blind

SixthSense is a wearable that helps blind and low-vision people sense obstacles around them and find a clear path. A phone is mounted on the chest and watches the way ahead. On-device models turn what the camera sees into a simple readout: how near obstacles are in the left, center, and right zones, what objects are present, and whether the path is clear. That readout drives a vibration belt worn at the waist, which buzzes on the side of the nearest obstacle so the user can feel which way to move. The point is that knowing something is close is not enough. A basic vibrating cane buzzes whenever anything is near, so in a crowd it buzzes constantly without telling you where the gap is. SixthSense reads each zone separately and steers the user toward open space, so it stays useful in busy areas. The user can also ask what is ahead and hear a short spoken answer, or point the camera at a sign and have its text read aloud. The vision runs on the phone. YOLOv11n detects objects and tags each to a left, center, or right zone. Depth-Anything-V2 estimates how near things are, which sets how hard the belt buzzes. Qwen2.5-0.5B answers spoken questions about the scene, and ML Kit reads text on demand. The models run through ExecuTorch as compiled files on the phone, offline, on a Qualcomm Snapdragon 8 Elite, with the option to run on the Hexagon NPU. The phone sends a small directional packet over Bluetooth to an ESP32, which drives the belt motors. Cost is the main reason it exists. Smart canes and glasses run from about $850 to over $2,000, and only one in ten people who need assistive technology can get it, dropping to about five percent in lower-income countries. SixthSense uses a phone the user already has and a sub-$20 belt, with room to reach about $50 at scale, putting this within reach of people who are priced out today.

Sub Rosa: On-Device AI Legal Intake

Sub Rosa is an Android app that turns a lawyer's first client meeting into clear intake notes. It runs entirely on the phone and never goes online. It records the conversation and transcribes it on the device with Whisper. A small language model, Qwen3-1.7B running with ExecuTorch on the Qualcomm Hexagon NPU, then produces three things: the facts, each shown next to the exact words from the transcript that back it up; the missing items, the standard intake points the client did not mention, found by comparing the talk against a checklist; and a few neutral follow-up questions to ask. The model only reads and sorts what it is given. It never states the law or gives advice. The legal knowledge lives in checklists written by people, not in the model's weights, so a small model stays accurate with no training step. Privacy is built in. The app has no internet permission, so the phone blocks any connection, and an on-screen panel shows nothing was sent. Audio is never saved, and all client data is encrypted on the device with AES-256-GCM.

HALO - Second Brain

HALO: Health Second Brain serves as a privacy-centric, on-device health companion specifically engineered to leverage the Snapdragon 8 Elite architecture of the Samsung Galaxy S25 Ultra. The application features an immersive and user-friendly interface built with Jetpack Compose, utilizing a high-contrast dark mode that highlights biometric data through dynamic organ-specific visualizations and real-time trend sparklines. At its core, HALO employs an advanced agentic framework that performs real-time intent classification to orchestrate interactions between local health data retrieval and automated web-based research tools. This system intelligently grounds every response in the user's personal biometrics—such as heart rate variability and sleep efficiency—which are pulled directly from a secure local SQLite database. These complex interactions are powered by a locally hosted Qwen-1.7B model, fine-tuned for health-specific logic and accelerated by the device's Hexagon NPU using the ExecuTorch and Qualcomm QNN stack. By maintaining all processing and inference strictly on-device, HALO provides a seamless, data-rich conversational experience while ensuring that sensitive biometric information never leaves the hardware.

PARLEY AI Orbital Negotiation System

PARLEY is a multi-agent orbital conjunction negotiation system built for the Band of Agents Hackathon 2026. It automates one of the most consequential decisions in space operations — who maneuvers, when, and by how much — without human intervention. The system runs six autonomous agents, each with a single responsibility: Sentinel monitors conjunction data and fires the first alert when collision probability crosses the safety threshold. Oracle enriches that alert with orbital parameters — fuel reserves, delta-V capacity, approach geometry — and recommends which satellite should maneuver. Operator Alpha and Operator Bravo represent each satellite operator. They negotiate directly, proposing burns, countering, and converging on a maneuver plan. Arbiter, running on an independent model, validates the agreed plan against coordination norms and issues the final verdict. Archivist seals every event into a hash-chained, tamper-proof audit trail — nothing can be altered after the fact. In a live test run, Sentinel flagged a critical conjunction between STARLINK-4412 and ONEWEB-2201 with a collision probability of 0.00018 (nearly double the 0.0001 threshold) and a miss distance of 42.3 meters. Oracle recommended ONEWEB-2201 as the maneuvering party based on its higher fuel reserves and delta-V capability. The operators negotiated and agreed on a 0.38 m/s retrograde burn, achieving a 500-meter miss distance and cutting collision probability by over 99%, to 1.4×10⁻⁶. Arbiter certified the plan, and Archivist sealed the full sequence into an immutable audit trail. PARLEY runs on Claude Sonnet 4.6 via the AI/ML API, with Featherless AI powering the independent Arbiter model, orchestrated through Band.ai's agent SDK. Every agent call is real, logged, and traceable. From detection to certified resolution — autonomously, in under a minute.

Agentic InferenceRoute

The traditional web services handle quick stateless requests measured in milliseconds, but LLM inference workloads require routing decisions that understand the current state of each model server — queue depth, memory utilization, loaded adapters, and ability to handle requests of different criticality levels. InferRoute adds a Featherless-powered classifier in front of that routing layer, and a band.ai cost-control agent watching the budget knob continuously.In Istio it fully conformant with GIE v1.0.1 and promotes GIE support to beta — separating platform admin concerns from AI/ML team concerns through InferencePool (platform admin resource) and InferenceObjective (AI/ML owner resource). InferRoute rides exactly that wave, adding a Featherless intelligence layer and band.ai budget governance on top of the freshest K8s-native inference primitive. It's not a toy demo — it maps directly to the production problem of routing requests based on KV-cache utilization, pending queue depth, and prefix affinity to minimize time-to-first-token, with cost control that enterprises actually need