Qwen

Qwen is a large language model family developed by the Qwen team at Alibaba Cloud. First released in 2023, the series spans dense and mixture-of-experts architectures across text, vision, and code, with most models published under the Apache 2.0 license. Developers can access Qwen models through Alibaba Cloud's Model Studio (DashScope) using an OpenAI-compatible API, or download weights directly from Hugging Face and GitHub.

General
Company	Qwen / Alibaba Cloud
Founded	2023 (first model release)
Headquarters	Hangzhou, China
Website	qwen.ai
Documentation	Qwen Docs
GitHub	github.com/QwenLM
Hugging Face	huggingface.co/Qwen
Type	LLM Provider / Open-Source AI Lab

Core Products

Qwen3 (Text LLMs)

Qwen3 is the flagship text model family, released in April 2025 under Apache 2.0. It includes dense models from 0.6B to 32B parameters and mixture-of-experts models up to 235B total parameters (22B active). All models support multilingual text generation, reasoning, tool use, and agentic workflows.

Qwen3-Coder

Qwen3-Coder is a coding-specialized model with 480B total parameters and 35B active, trained on 7.5 trillion tokens with a 70% code-focused dataset. Released in July 2025, it achieves state-of-the-art results among open models on SWE-Bench Verified.

Qwen3.6 (Vision-Language)

Qwen3.6 is a multimodal model with a unified vision-language architecture trained on trillions of multimodal tokens. It supports text and image inputs across 201 languages and dialects, with capabilities covering reasoning, coding, and visual understanding.

Qwen-Image-2.0

Qwen-Image-2.0 is a 7B-parameter image generation model supporting photorealism, professional typography, and unified generation-editing workflows, released in February 2026.

Qwen-MT

Qwen-MT is a translation model covering 92 major languages and dialects, reaching over 95% of the global population. It is designed for high-quality translation in production pipelines.

Qwen Code

Qwen Code is an open-source terminal coding agent optimized for the Qwen model series. It supports writing features, fixing bugs, navigating large codebases, and generating pull requests, with GitHub Actions integration available.

Developer Resources

Qwen models are accessible through Alibaba Cloud Model Studio (DashScope) via an OpenAI-compatible API, or as open weights on Hugging Face. The API supports both text-only and multimodal models.

Helpful Links

Qwen Documentation (model cards, quickstarts, and research notes)
Alibaba Cloud Model Studio (API reference and DashScope SDK docs)
GitHub (QwenLM) (open-source model weights, code, and tools)
Hugging Face (Qwen) (download model weights and datasets)
Qwen API Platform (get an API key and start building)
Qwen Studio (web interface for chat, image understanding, and generation)

Key Features

Open weights under Apache 2.0 Most Qwen3 models are released under Apache 2.0, allowing commercial use, fine-tuning, and redistribution without restrictions.

OpenAI-compatible API Qwen models are served through DashScope using the OpenAI-compatible endpoint format, making it straightforward to use Qwen models in existing OpenAI SDK integrations.

Multilingual coverage Qwen3.6 supports 201 languages and dialects. Qwen-MT covers 92 major languages for dedicated translation tasks.

Mixture-of-Experts (MoE) architecture The largest Qwen3 models use MoE, activating only a subset of total parameters per token (for example, 22B of 235B active). This reduces inference cost relative to comparably capable dense models.

Use Cases

Agentic coding workflows Qwen3-Coder and Qwen Code are designed for software development tasks: writing features, fixing bugs, navigating large codebases, and generating pull requests via the terminal or CI pipelines.

Multilingual applications Qwen-MT and Qwen3.6's broad language support make them suitable for translation tools, multilingual chatbots, and localized content pipelines.

Multimodal document and image processing Qwen3.6 handles image understanding, document analysis, and visual reasoning alongside text, enabling applications like document Q&A and visual search.

Trajectory — Decision-Guided Delivery

A customer cancels a subscription renewal and unexpectedly loses access to time they already paid for. Trajectory turns that friction into a decision the people responsible for the experience can make without writing code. A guided GitHub issue captures what happened and the desired result. The installed Trajectory GitHub App presents clear outcomes: keep paid access, end it immediately, or propose a prorated refund. Selecting an outcome prepares a corresponding code and configuration change in a reviewable branch and pull request. Changing the decision creates a separate outcome and invalidates the earlier intent confirmation. A Vultr-hosted Go worker uses Vultr Serverless Inference to generate a bounded access-function candidate. A disposable, isolated container runs independent boundary tests; repository-defined assertions and a comparison against the proposed PR tree check the selected change. GitHub Checks connect the issue, choice, patch, test results and evidence for review. A separate containment demonstration tests that a host canary is inaccessible from the sandbox. The demo uses an illustrative customer and a fixed example proposing $18 for 18 unused days. It does not execute a refund or change a real customer account. Verification covers the stated test cases; intent review remains a human step before merge.

ROCm Plate & Sign OCR (AMD AI League)

An OCR container for AMD AI League Match 2 that reads licence plates and road signs with an open vision-language model running on an AMD GPU through ROCm. How it works: the container starts a resident server that loads Qwen2.5-VL-7B-Instruct once, in bf16 on the GPU, runs a warm-up inference and then listens on a unix socket. Each call to app.py is a thin client, so the harness can start a new process per image without the model ever being reloaded. Image handling accepts PNG, JPEG and TIFF (including 16-bit and multi-page files), applies EXIF rotation, and resizes very large or very small images to keep the number of visual tokens, and the time per image, bounded. The prompt encodes the transcription rules: return the plate number without state names or slogans, keep the province character and letter on Chinese plates, keep every word printed on a sign, and join multi-line text top to bottom. Post-processing removes labels, quotes and explanations, and folds look-alike separators and full-width characters that the grader's normalization would not remove. Reliability: the model weights are baked into the image, so it runs with no network access. Every call writes a valid JSON file, even when something fails. The build checks that torch and torchvision are ROCm builds. The image is built and contract-checked in GitHub Actions, and evaluated against the sample images plus blurred, noisy, glare, low-light and skewed variants.

Aperture - Web Retrieval with Traceable Evidence

Aperture is an AI web retrieval agent built for AMD AI League Match 4. It accepts a natural-language question and an optional starting URL, navigates web pages, and returns a concise answer with supporting source URLs. The system combines asynchronous retrieval, structure-preserving parsing, and a locally hosted Qwen3-4B-Instruct-2507 model. It preserves headings, tables, and labeled fields so that names, dates, versions, and numeric values retain their context. Evidence checks verify literal support, source provenance, answer types, and requested qualifiers. When available evidence is insufficient, the agent returns an empty answer. Inference runs locally, and retrieval uses no paid search service or remote inference API. The submission is packaged as a prebuilt Docker image using the required ROCm base, with model weights included. Local validation covers extraction, output formatting, runtime behavior, and factual retrieval. General web accuracy remains limited by search availability, page access, model errors, and time budgets. AMD GPU execution and official evaluator acceptance have not yet been verified. The public repository includes implementation details, benchmark reports, and reproducibility documentation.

Exact Web Answer Agent (AMD AI League Match 4)

Match 4 entry: an agent that reads a natural-language request, navigates the given website (or a saved copy of it) and returns the one specific value the user asked for, with the page it came from. Not a summary and not a list of links. Match 4 image: docker.io/driedsandwich/amd-ai-league:m4-v1 How it works. Each request runs as a fresh python3 /app/app.py process with a 24-second internal deadline. It writes a well-formed placeholder JSON first, then fetches the start page with short timeouts, follows the same-site links whose text best matches the question, cleans each page to text (tables kept as rows) and ranks passages. A Qwen3-4B-Instruct model, loaded once at container start and kept resident on the AMD GPU behind a unix socket, picks the value from the top passages and names the supporting page. A rule-based extractor for prices, dates, versions, numbers and contacts is the fallback when the model has no time left. The answer is the value only; if the pages do not contain it, the agent returns an empty answer instead of guessing. Output: /app/output/<query-id>_output.json with answer, citations (source URLs, or corpus-relative paths when pages are given as a folder), sources, source_url and confidence. Because the detailed Match 4 harness was not published, the image follows the container contract shared by the earlier Matches: the mandated ROCm PyTorch base image, weights shipped inside the image, no downloads at start-up, and the model placed on the GPU explicitly. The script accepts several argument spellings (--query or --question, --query-id or --id, --url, --corpus for saved pages, --index, --input). No API keys; inference runs locally on the AMD GPU. Match 3 and Match 2 images are given in their Mini Challenge fields (docker.io/driedsandwich/amd-ai-league:mc3-v1 and docker.io/driedsandwich/amd-ai-league:mc2-v2).

Exact Web Answer Agent (AMD AI League Match 4)

AMD ROCm Grounded Web Agent

GroundedROCm is a 100% ROCm-native grounded retrieval engine powered by Qwen2.5-0.5B on AMD MI300X delivering verified JSON answers with zero hallucinations. Traditional search suffers from CUDA dependencies, high latency, API costs, and hallucinations. We solve it with a 5-stage local pipeline: Planner (Qwen2.5-0.5B via http://llama.cpp with fast-path for simple queries), Search Layer (Tavily / Qdrant / JSONL pluggable), Evidence Engine (deduplication + BM25 + compression), Synthesizer (warm http://llama.cpp server with schema-constrained JSON, max 96 tokens), and Guardrails (JSON repair + number/name grounding + URL validation). Result: 15MB binary via http://llama.cpp, small Docker size, ∼0.6s latency on MI300X, 98% grounding accuracy, 100% source validity, zero CUDA, zero external cost, instantly switchable to internal Qdrant for enterprise deployment.