SYLIN

Microsoft

Maker · 21 entries

Everything from Microsoft

NameWhat to knowPrice or memoryRanks
MAI ImageModelmixed Strong at image editing and cheap, but trails GPT Image Sunburst on Arena#2 image editing#3 image generation
Aion-1.0-Instruct-2.3BModelNew open model on Hugging Face (text-generation): 2.29B parameters, 4.58 GB of weights, license mit, 4 community quantizations incl. GGUF~1.4 GB
Copilot FusionApp or serviceMulti-model harness reports 71.2% on Terminal-Bench 4.1 at 88% lower cost than Claude Fable 5.2
MAI CodeModelunproven Microsoft says it nearly matches top coding models cheaply; no independent tests yet
MAI Code FlashModelunproven Microsoft says it matches cheaper rivals on coding, but no one else has tested it
MAI Cyber FlashModelMAI Cyber 2 Flash for agentic codebase defense; output tokens 86% cheaper than Cyber 1.1 Flash; existing MDASH customers can move to it
MAI DecisionModelDecision 1.1 routing model is generally available, used by 4,000 customers
MAI Image FlashModelunproven Cheap, fast image generation at about 1 cent each; quality not yet independently tested
MAI Licensed WeightsApp or serviceunproven Announced Azure-only customization of MAI models; not downloadable, no price or date
MAI RealtimeModelMAI Realtime 1, Microsoft's first live speech-to-speech model, starts at $0.01 per audio minute; try it in MAI Labs
MAI TranscribeModelunproven Microsoft says it is the fastest and most accurate; no independent tests yet
VibeVoice-ASRModelSupports over 50 languages natively; finetuning code and vLLM inference are available~1.6 GB
VibeVoice-TTSModelpromising Small, MIT-licensed speech model from Microsoft, but meant for research only~1.6 GB
QuicksandToolMicrosoft's Quicksand library bundles QEMU and Alpine Linux in a Python package; used by external-sandbox-quicksand
PlaywrightFrameworkstrong Established browser automation framework that newer tools are compared against and still preferred
Lazarus (Microsoft Research weight-tampering study)PaperA small weight change can leave an agent normal in routine tests but altered under specific conditions; advice: pin model versions, verify signatures and hashes
MoGe-3ModelUsed as depth baseline that Marigold v2 outperforms
VibeVoice-ASR-StreamingModelUnified streaming ASR continuously transcribes who said what as speech arrives, with custom hotwords and 10 languages
FaraModelFara1.5 released in 4B, 9B and 27B sizes, a much improved successor to the earlier Fara-7B web computer-use agent~76 GB
Mage FlowModelCompact 4B-parameter model for text-to-image generation and instruction-based image editing, runnable locally on one GPU~17.5 GB
VibeVoice-RealtimeModelAdded experimental speakers: multilingual voices in nine languages and 11 English style voices