// Open-model VRAM table · family
Microsoft VibeVoice / Fara / Harrier: local deployment & VRAM
12 open-weight models from Microsoft (14 official repos incl. quantised builds). At 4-bit they range from 3.8 GB (Fara1.5-4B) to 19.3 GB (Fara1.5-27B).
official orgs: microsoftlatest 2026-09-02memory = estimate
01Vision-language (4)
| Model | Params | Context | BF16 | 4-bit | Smallest GPU (4-bit) | Mac (4-bit) | Licence | Released |
|---|---|---|---|---|---|---|---|---|
| Fara1.5-4B ↗ | 4.54B | 256K | 10.8 GB | 3.8 GB | 8GB GPU | Mac 16GB | Commercial OK | 2026-07-17 |
| Fara1.5-27B ↗ | 27.4B | 256K | 61.2 GB | 19.3 GB | RTX 4090 24GB | Mac 36GB | Commercial OK | 2026-07-17 |
| Fara1.5-9B | 9.41B | 256K | 21.5 GB | 7.0 GB | 8GB GPU | Mac 16GB | Commercial OK | 2026-05-12 |
| Fara-7B | 8.29B | 125K | 19.2 GB | 6.5 GB | 8GB GPU | Mac 16GB | Commercial OK | 2025-10-30 |
02Embedding (3)
| Model | Params | BF16 | 4-bit | Smallest GPU (4-bit) | Mac (4-bit) | Licence | Released |
|---|---|---|---|---|---|---|---|
| harrier-oss-v1-270m | 268M | 1.1 GB | 0.7 GB | 8GB GPU | Mac 16GB | Commercial OK | 2026-03-30 |
| harrier-oss-v1-0.6b | 596M | 1.8 GB | 0.9 GB | 8GB GPU | Mac 16GB | Commercial OK | 2026-03-30 |
| harrier-oss-v1-27b | 27B | 59.9 GB | 18.5 GB | RTX 4090 24GB | Mac 36GB | Commercial OK | 2026-03-30 |
03Speech-to-text (3)
| Model | Params | BF16 | 4-bit | Smallest GPU (4-bit) | Mac (4-bit) | Licence | Released |
|---|---|---|---|---|---|---|---|
| VibeVoice-ASR-Streaming-1.5B | 2.81B | 6.7 GB | 2.4 GB | 8GB GPU | Mac 16GB | Commercial OK | 2026-09-02 |
| VibeVoice-ASR-Streaming-7B | 8.67B | 19.6 GB | 6.3 GB | 8GB GPU | Mac 16GB | Commercial OK | 2026-09-02 |
| VibeVoice-ASR | 8.67B | 19.6 GB | 6.3 GB | 8GB GPU | Mac 16GB | Commercial OK | 2026-01-21 |
04Text-to-speech / audio (2)
| Model | Params | BF16 | 4-bit | Smallest GPU (4-bit) | Mac (4-bit) | Licence | Released |
|---|---|---|---|---|---|---|---|
| VibeVoice-Realtime-0.5B | 1.02B | 2.7 GB | 1.2 GB | 8GB GPU | Mac 16GB | Commercial OK | 2025-12-04 |
| VibeVoice-1.5B | 2.7B | 6.4 GB | 2.3 GB | 8GB GPU | Mac 16GB | Commercial OK | 2025-08-25 |
05Other families
Qwen · IBM Granite · NVIDIA Nemotron · Gemma · Mistral / Magistral / Devstral · GLM / CogVideoX (Z.ai) · DeepSeek · Ai2 OLMo / Molmo · Ant Ling / Ring / Ming · Liquid LFM · MiniCPM / VoxCPM · Tencent Hunyuan / Hy · Phi · Llama · Meituan LongCat · Jina embeddings & rerankers · Wan · Kimi · Cohere Command / Aya · Xiaomi MiMo · FLUX · InternLM / Intern-S · StepFun Step · MiniMax · Baidu ERNIE · NVIDIA Parakeet / Canary speech · Stable Diffusion / Stable Audio · Yi (01.AI) · ByteDance Seed · SmolLM / SmolVLM · BGE embeddings & rerankers · LTX Video · FunAudioLLM speech · gpt-oss · poolside Laguna · Whisper · Baichuan · Thinking Machines Inkling · xAI Grok · rednote dots
Totals include 10% + 0.5GB overhead and one 8K-token KV cache for text models. Estimates only. Method →