model monday
2 TopicsNow in Foundry: Qwen3-Coder-Next, Qwen3-ASR-1.7B, Z-Image
This week's spotlight features three models from that demonstrate enterprise-grade AI across the full scope of modalities. From low latency coding agents to state-of-the-art multilingual speech recognition and foundation-quality image generation, these models showcase the breadth of innovation happening in open-source AI. Each model balances performance with practical deployment considerations, making them viable for production systems while pushing the boundaries of what's possible in their respective domains. This week's Model Mondays edition highlights Qwen3-Coder-Next, an 80B MoE model that activates only 3B parameters while delivering coding agent capabilities with 256k context; Qwen3-ASR-1.7B, which achieves state-of-the-art accuracy across 52 languages and dialects; and Z-Image from Tongyi-MAI, an undistilled text-to-image foundation model with full Classifier-Free Guidance support for professional creative workflows. Models of the week Qwen: Qwen3-Coder-Next Model Specs Parameters / size: 80B total (3B activated) Context length: 262,144 tokens Primary task: Text generation (coding agents, tool use) Why it's interesting Extreme efficiency: Activates only 3B of 80B parameters while delivering performance comparable to models with 10-20x more active parameters, making advanced coding agents viable for local deployment on consumer hardware Built for agentic workflows: Excels at long-horizon reasoning, complex tool usage, and recovering from execution failures, a critical capability for autonomous development that go beyond simple code completion Benchmarks: Competitive performance with significantly larger models on SWE-bench and coding benchmarks (Technical Report) Try it Use Case Prompt Pattern Code generation with tool use Provide task context, available tools, and execution environment details Long-context refactoring Include full codebase context within 256k window with specific refactoring goals Autonomous debugging Present error logs, stack traces, and relevant code with failure recovery instructions Multi-file code synthesis Describe architecture requirements and file structure expectations Financial services sample prompt: You are a coding agent for a fintech platform. Implement a transaction reconciliation service that processes batches of transactions, detects discrepancies between internal records and bank statements, and generates audit reports. Use the provided database connection tool, logging utility, and alert system. Handle edge cases including partial matches, timing differences, and duplicate transactions. Include unit tests with 90%+ coverage. Qwen: Qwen3-ASR-1.7B Model Specs Parameters / size: 1.7B Context length: 256 tokens (default), configurable up to 4096 Primary task: Automatic speech recognition (multilingual) Why it's interesting All-in-one multilingual capability: Single 1.7B model handles language identification plus speech recognition for 30 languages, 22 Chinese dialects, and English accents from multiple regions—eliminating the need to manage separate models per language Specialized audio versatility: Transcribes not just clean speech but singing voice, songs with background music, and extended audio files, expanding use cases beyond traditional ASR to entertainment and media workflows State-of-the-art accuracy: Outperforms GPT-4o, Gemini-2.5, and Whisper-large-v3 across multiple benchmarks. English: Tedlium 4.50 WER vs 7.69/6.15/6.84; Chinese: WenetSpeech 4.97/5.88 WER vs 15.30/14.43/9.86 (Technical Paper) Language ID included: 97.9% average accuracy across benchmark datasets for automatic language identification, eliminating the need for separate language detection pipelines Try it Use Case Prompt Pattern Multilingual transcription Send audio files via API with automatic language detection Call center analytics Process customer service recordings to extract transcripts and identify languages Content moderation Transcribe user-generated audio content across multiple languages Meeting transcription Convert multilingual meeting recordings to text for documentation Customer support sample prompt: Deploy Qwen3-ASR-1.7B to a Microsoft Foundry endpoint and transcribe multilingual customer service calls. Send audio files via API to automatically detect the language (from 52 supported options including 30 languages and 22 Chinese dialects) and generate accurate transcripts. Process calls from customers speaking English, Spanish, Mandarin, Cantonese, Arabic, French, and other languages without managing separate models per language. Use transcripts for quality assurance, compliance monitoring, and customer sentiment analysis. Tongyi-MAI: Z-Image Model Specs Parameters / size: 6B Context length: N/A (text-to-image) Primary task: Text-to-image generation Why it's interesting Undistilled foundation model: Full-capacity base without distillation preserves complete training signal with Classifier-Free Guidance support (a technique that improves prompt adherence and output quality), enabling complex prompt engineering and negative prompting that distilled models cannot achieve High output diversity: Generates distinct character identities in multi-person scenes with varied compositions, facial features, and lighting, critical for creative applications requiring visual variety rather than consistency Aesthetic versatility: Handles diverse visual styles from hyper-realistic photography to anime and stylized illustrations within a single model, supporting resolutions from 512×512 to 2048×2048 at any aspect ratio with 28-50 inference steps (Technical Paper) Try it Use Case Prompt Pattern Multilingual transcription Send audio files via API with automatic language detection Call center analytics Process customer service recordings to extract transcripts and identify languages Content moderation Transcribe user-generated audio content across multiple languages Meeting transcription Convert multilingual meeting recordings to text for documentation E-commerce sample prompt: Professional product photography of a modern ergonomic office chair in a bright Scandinavian-style home office. Natural window lighting from left, clean white desk with laptop and succulent plant, light oak hardwood floor. Chair positioned at 45-degree angle showing design details. Photorealistic, commercial photography, sharp focus, 85mm lens, f/2.8, soft shadows. Getting started You can deploy open‑source Hugging Face models directly in Microsoft Foundry by browsing the Hugging Face collection in the Foundry model catalog and deploying to managed endpoints in just a few clicks. You can also start from the Hugging Face Hub. First, select any supported model and then choose "Deploy on Microsoft Foundry", which brings you straight into Azure with secure, scalable inference already configured. Learn how to discover models and deploy them using Microsoft Foundry documentation. Follow along the Model Mondays series and access the GitHub to stay up to date on the latest Read Hugging Face on Azure docs Learn about one-click deployments from the Hugging Face Hub on Microsoft Foundry Explore models in Microsoft Foundry374Views0likes0CommentsWhat is trending in Hugging Face on Microsoft Foundry? Feb, 2, 2026
Open‑source AI is moving fast, with important breakthroughs in reasoning, agentic systems, multimodality, and efficiency emerging every day. Hugging Face has been a leading platform where researchers, startups, and developers share and discover new models. Microsoft Foundry brings these trending Hugging Face models into a production‑ready experience, where developers can explore, evaluate, and deploy them within their Azure environment. Our weekly Model Monday’s series highlights Hugging Face models available in Foundry, focusing on what matters most to developers: why a model is interesting, where it fits, and how to put it to work quickly. This week’s Model Mondays edition highlights three Hugging Face models, including a powerful Mixture-of-Experts model from Z. AI designed for lightweight deployment, Meta’s unified foundation model for image and video segmentation, and MiniMax’s latest open-source agentic model optimized for complex workflows. Models of the week Z.AI’s GLM-4.7-flash Model Basics Model name: zai-org/GLM-4.7-Flash Parameters / size: 30B total -3B active Default settings: 131,072 max new tokens Primary task: Agentic, Reasoning and Coding Why this model matters Why it’s interesting: It utilizes a Mixture-of-Experts (MoE) architecture (30B total parameters and 3B active parameters) to offer a new option for lightweight deployment. It demonstrates strong performance on logic and reasoning benchmarks, outperforming similar sized models like gpt-oss-20b on AIME 25 and GPQA benchmarks. It supports advanced inference features like "Preserved Thinking" mode for multi-turn agentic tasks. Best‑fit use cases: Lightweight local deployment, multi-turn agentic tasks, and logical reasoning applications. What’s notable: From the Foundry catalog, users can deploy on a A100 instance or unsloth/GLM-4.7-Flash-GGUF on a CPU. ource SOTA scores among models of comparable size. Additionally, compared to similarly sized models, GLM-4.7-Flash demonstrates superior frontend and backend development capabilities. Click to see more: https://docs.z.ai Try it Use case Best‑practice prompt pattern Agentic coding (multi‑step repo work, debugging, refactoring) Treat the model as an autonomous coding agent, not a snippet generator. Explicitly require task decomposition and step‑by‑step execution, then a single consolidated result. Long‑context agent workflows (local or low‑cost autonomous agents) Call out long‑horizon consistency and context preservation. Instruct the model to retain earlier assumptions and decisions across turns. Now that you know GLM‑4.7‑Flash works best when you give it a clear goal and let it reason through a bounded task, here’s an example prompt that a product or engineering team might use to identify risks and propose mitigations: You are a software reliability analyst for a mid‑scale SaaS platform. Review recent incident reports, production logs, and customer issues to uncover edge‑case failures outside normal usage (e.g., rare inputs, boundary conditions, timing/concurrency issues, config drift, or unexpected feature interactions). Prioritize low‑frequency, high‑impact risks that standard testing misses. Recommend minimal, low‑cost fixes (validation, guardrails, fallback logic, or documentation). Deliver a concise executive summary with sections: Observed Edge Cases, Root Causes, User Impact, Recommended Lightweight Fixes, and Validation Steps. Meta's Segment Anything 3 (SAM3) Model Basics Model name: facebook/sam3 Parameters / size: 0.9B Primary task: Mask Generation, Promptable Concept Segmentation (PCS) Why this model matters Why it’s interesting: It handles a vastly larger set of open-vocabulary prompts than SAM 2, and unifies image and video segmentation capabilities. It includes a "SAM 3 Tracker" mode that acts as a drop-in replacement for SAM 2 workflows with improved performance. Best‑fit use cases: Open-vocabulary object detection, video object tracking, and automatic mask generation What’s notable: Introduces Promptable Concept Segmentation (PCS), allowing users to find all matching objects (e.g., "dial") via text prompt rather than just single instances. Try it This model enables users to identify specific objects within video footage and isolate them over extended periods. With just one line of code, it is possible to detect multiple similar objects simultaneously. The accompanying GIF demonstrates how SAM3 efficiently highlights players wearing white on the field as they appear and disappear from view. Additional examples are available at the following repository: https://github.com/facebookresearch/sam3/blob/main/assets/player.gif Use case Best‑practice prompt pattern Agentic coding (multi‑step repo work, debugging, refactoring) Treat SAM 3 as a concept detector, not an interactive click tool. Use short, concrete noun‑phrase concept prompts instead of describing the scene or asking questions. Example prompt: “yellow school bus” or “shipping containers”. Avoid verbs or full sentences. Video segmentation + object tracking Specify the same concept prompt once, then apply it across the video sequence. Do not restate the prompt per frame. Let the model maintain identity continuity. Example: “person wearing a red jersey”. Hard‑to‑name or visually subtle objects Use exemplar‑based prompts (image region or box) when text alone is ambiguous. Optionally combine positive and negative exemplars to refine the concept. Avoid over‑constraining with long descriptions. Using the GIF above as a leading example, here is a prompt that shows how SAM 3 turns raw sports footage into structured, reusable data. By identifying and tracking players based on visual concepts like jersey color so that sports leagues can turn tracked data into interactive experiences where automated player identification can relay stats, fun facts, etc when built into a larger application. Here is a prompt that will allow you to start identifying specific players across video: Act as a sports analytics operator analyzing football match footage. Segment and track all football players wearing blue jerseys across the video. Generate pixel‑accurate segmentation masks for each player and assign persistent instance IDs that remain stable during camera movement, zoom, and player occlusion. Exclude referees, opposing team jerseys, sidelines, and crowd. Output frame‑level masks and tracking metadata suitable for overlays, player statistics, and downstream analytics pipelines. MiniMax AI's MiniMax-M2.1 Model Basics Model name: MiniMaxAI/MiniMax-M2.1 Parameters / size: 229B-10B Active Default settings: 200,000 max new tokens Primary task: Agentic and Coding Why this model matters Why it’s interesting: It is optimized for robustness in coding, tool use, and long-horizon planning, outperforming Claude Sonnet 4.5 in multilingual scenarios. It excels in full-stack application development, capable of architecting apps "from zero to one”. Previous coding models focused on Python optimization, M2.1 brings enhanced capabilities in Rust, Java, Golang, C++, Kotlin, Objective-C, TypeScript, JavaScript, and other languages. The model delivers exceptional stability across various coding agent frameworks. Best‑fit use cases: Lightweight local deployment, multi-turn agentic tasks, and logical reasoning applications. What’s notable: The release of open-source weights for M2.1 delivers a massive leap over M2 on software engineering leaderboards. https://www.minimax.io/ Try it Use case Best‑practice prompt pattern End‑to‑end agentic coding (multi‑file edits, run‑fix loops) Treat the model as an autonomous coding agent, not a snippet generator. Explicitly require task decomposition and step‑by‑step execution, then a single consolidated result. Long‑horizon tool‑using agents (shell, browser, Python) Explicitly request stepwise planning and sequential tool use. M2.1’s interleaved thinking and improved instruction‑constraint handling are designed for complex, multi‑step analytical tasks that require evidence tracking and coherent synthesis, not conversational back‑and‑forth. Long‑context reasoning & analysis (large documents / logs) Declare the scope and desired output structure up front. MiniMax‑M2.1 performs best when the objective and final artifact are clear, allowing it to manage long context and maintain coherence. Because MiniMax‑M2.1 is designed to act as a long‑horizon analytical agent, it shines when you give it a clear end goal and let it work through large volumes of information—here’s a prompt a risk or compliance team could use in practice: You are a financial risk analysis agent. Analyze the following transaction logs and compliance policy documents to identify potential regulatory violations and systemic risk patterns. Plan your approach before executing. Work through the data step by step, referencing evidence where relevant. Deliver a final report with the following sections: Key Risk Patterns Identified, Supporting Evidence, Potential Regulatory Impact, Recommended Mitigations. Your response should be a complete, executive-ready report, not a conversational draft. Getting started You can deploy open‑source Hugging Face models directly in Microsoft Foundry by browsing the Hugging Face collection in the Foundry model catalog and deploying to managed endpoints in just a few clicks. You can also start from the Hugging Face Hub. First, select any supported model and then choose "Deploy on Microsoft Foundry", which brings you straight into Azure with secure, scalable inference already configured. Learn how to discover models and deploy them using Microsoft Foundry documentation. Follow along the Model Mondays series and access the GitHub to stay up to date on the latest Read Hugging Face on Azure docs Learn about one-click deployments from the Hugging Face Hub on Microsoft Foundry Explore models in Microsoft Foundry642Views0likes0Comments