Mention GGUF, sort models by downloads, list bf16/8-bit first

#2
by Paris99 - opened
Files changed (1) hide show
  1. README.md +11 -10
README.md CHANGED
@@ -11,7 +11,8 @@ pinned: false
11
 
12
  **Open models, ready to run on Apple Silicon.**
13
 
14
- We convert and test open models for [MLX](https://github.com/ml-explore/mlx),
 
15
  so they run locally on a Mac with no cloud and no API keys. We check every
16
  build against the original model before it goes up, and we publish speed
17
  numbers from real hardware (MacBook Pro M5 Max, 128 GB).
@@ -20,19 +21,19 @@ numbers from real hardware (MacBook Pro M5 Max, 128 GB).
20
 
21
  | Family | Builds | What it's for |
22
  |---|---|---|
23
- | **Fastino GLiDE 27B** | [8-bit](https://huggingface.co/netprojects/GLiDE-mlx-8bit) · [4-bit](https://huggingface.co/netprojects/GLiDE-mlx-4bit) | Calibrated decision model (choice / yes-no / score), midloop, full-attention reading |
24
- | **H2O-Lightning 4B** | [q8](https://huggingface.co/netprojects/h2o-lightning-4b-mlx-q8) · [q4](https://huggingface.co/netprojects/h2o-lightning-4b-mlx-q4) | Small calibrated decision model from H2O.ai (text and images, two-way reading, Box questions) |
25
- | **IBM Granite Switch 4.1** | 8B [8-bit](https://huggingface.co/netprojects/granite-switch-4.1-8b-mlx-8bit) · [4-bit](https://huggingface.co/netprojects/granite-switch-4.1-8b-mlx-4bit) — 30B [8-bit](https://huggingface.co/netprojects/granite-switch-4.1-30b-mlx-8bit) · [4-bit](https://huggingface.co/netprojects/granite-switch-4.1-30b-mlx-4bit) | IBM Granite with switchable chat, decision and guardian modes |
26
- | **NVIDIA Nemotron Dory 1.0 4B** | [8-bit](https://huggingface.co/netprojects/nemotron-dory1.0-4b-mlx-8bit) · [bf16](https://huggingface.co/netprojects/nemotron-dory1.0-4b-mlx-bf16) · [NVFP4 4-bit](https://huggingface.co/netprojects/nemotron-dory1.0-4b-mlx-nvfp4) | Looped reasoning LM from NVIDIA Nemotron Labs |
27
- | **SenseNova-Loop-9B-Preview** | [8-bit](https://huggingface.co/netprojects/SenseNova-Loop-9B-Preview-MLX-8bit) · [6-bit](https://huggingface.co/netprojects/SenseNova-Loop-9B-Preview-MLX-6bit) · [4-bit](https://huggingface.co/netprojects/SenseNova-Loop-9B-Preview-MLX-4bit) | Looped Qwen3.5-based LM from SenseNova (layers 4–27 run twice, two-stream mHC) |
28
- | **NVIDIA Cosmos3-Edge reasoner** | [8-bit](https://huggingface.co/netprojects/Cosmos3-Edge-mlx-8bit) · [6-bit](https://huggingface.co/netprojects/Cosmos3-Edge-mlx-6bit) | Vision-language reasoner from NVIDIA: text, image and video in, text out |
29
- | **NVIDIA Cosmos3-Edge video generator** | [8-bit](https://huggingface.co/netprojects/Cosmos3-Edge-Omni-mlx-8bit) · [bf16](https://huggingface.co/netprojects/Cosmos3-Edge-Omni-mlx-bf16) | Video generator from NVIDIA: text or a first frame in, video out (text-to-video, image-to-video, text-to-image) |
30
  | **Tencent DSpine drafter for Qwen3.8-27B** | [drafter](https://huggingface.co/netprojects/Qwen3.8-27B-DSpine-b16-MLX) | Speculative decoding for Qwen3.8-27B: same output, 2.0x faster on 4-bit, 2.9x on 8-bit, 4.3x on bf16, and 2x or more with 33k-token prompts |
31
  | **NVIDIA Nemotron 3.5 Embed / Rerank** | Embed-1B [bf16](https://huggingface.co/netprojects/Nemotron-3.5-Embed-1B-mlx-bf16) · [8-bit](https://huggingface.co/netprojects/Nemotron-3.5-Embed-1B-mlx-8bit) — Embed-8B [8-bit](https://huggingface.co/netprojects/Nemotron-3.5-Embed-8B-mlx-8bit) · [4-bit](https://huggingface.co/netprojects/Nemotron-3.5-Embed-8B-mlx-4bit) — Rerank-8B [8-bit](https://huggingface.co/netprojects/Nemotron-3.5-Rerank-8B-mlx-8bit) · [4-bit](https://huggingface.co/netprojects/Nemotron-3.5-Rerank-8B-mlx-4bit) · [GGUF](https://huggingface.co/netprojects/Nemotron-3.5-Rerank-8B-GGUF) | Retrieval: embeddings and reranking for text and page images (scores match NVIDIA's published ViDoRe v3 results) |
 
 
 
32
  | **AutoTrust JEV-35B** | [8-bit](https://huggingface.co/netprojects/AutoTrust-JEV-35B-mlx-8bit) · [4-bit mixed](https://huggingface.co/netprojects/AutoTrust-JEV-35B-mlx-4bit-mixed) · [3-bit mixed](https://huggingface.co/netprojects/AutoTrust-JEV-35B-mlx-3bit-mixed) | AutoTrust decision and classification model |
33
  | **AutoTrust-JEV-27B v1.1** | [8-bit](https://huggingface.co/netprojects/AutoTrust-JEV-27B-v1.1-mlx-8bit) · [4-bit](https://huggingface.co/netprojects/AutoTrust-JEV-27B-v1.1-mlx-4bit) | Decision model with bidirectional prompt reading (custom loader included) |
34
- | **Qwen Concept 27B** | [6-bit](https://huggingface.co/netprojects/Qwen-Concept-27B-MLX-6bit) · [8-bit](https://huggingface.co/netprojects/Qwen-Concept-27B-MLX-8bit) · [4-bit](https://huggingface.co/netprojects/Qwen-Concept-27B-MLX-4bit) · [bf16](https://huggingface.co/netprojects/Qwen-Concept-27B-MLX-bf16) | Qwen3.8-27B edited by Concept Country to reach answers with less thinking; text and images |
35
- | **Laplace-1-Atlas** | [6-bit](https://huggingface.co/netprojects/Laplace-1-Atlas-MLX-6bit) · [8-bit](https://huggingface.co/netprojects/Laplace-1-Atlas-MLX-8bit) · [4-bit](https://huggingface.co/netprojects/Laplace-1-Atlas-MLX-4bit) · [bf16](https://huggingface.co/netprojects/Laplace-1-Atlas-MLX-bf16) · [GGUF](https://huggingface.co/netprojects/Laplace-1-Atlas-GGUF) | Juneo Lab's 27B decision model: scores candidates in one forward pass; JevBench 91.8% (6-bit) |
 
 
 
36
 
37
  ## How we build
38
 
 
11
 
12
  **Open models, ready to run on Apple Silicon.**
13
 
14
+ We convert and test open models for [MLX](https://github.com/ml-explore/mlx)
15
+ (and GGUF for llama.cpp),
16
  so they run locally on a Mac with no cloud and no API keys. We check every
17
  build against the original model before it goes up, and we publish speed
18
  numbers from real hardware (MacBook Pro M5 Max, 128 GB).
 
21
 
22
  | Family | Builds | What it's for |
23
  |---|---|---|
24
+ | **NVIDIA Nemotron Dory 1.0 4B** | [bf16](https://huggingface.co/netprojects/nemotron-dory1.0-4b-mlx-bf16) · [8-bit](https://huggingface.co/netprojects/nemotron-dory1.0-4b-mlx-8bit) · [NVFP4 4-bit](https://huggingface.co/netprojects/nemotron-dory1.0-4b-mlx-nvfp4) | Looped reasoning LM from NVIDIA Nemotron Labs |
 
 
 
 
 
 
25
  | **Tencent DSpine drafter for Qwen3.8-27B** | [drafter](https://huggingface.co/netprojects/Qwen3.8-27B-DSpine-b16-MLX) | Speculative decoding for Qwen3.8-27B: same output, 2.0x faster on 4-bit, 2.9x on 8-bit, 4.3x on bf16, and 2x or more with 33k-token prompts |
26
  | **NVIDIA Nemotron 3.5 Embed / Rerank** | Embed-1B [bf16](https://huggingface.co/netprojects/Nemotron-3.5-Embed-1B-mlx-bf16) · [8-bit](https://huggingface.co/netprojects/Nemotron-3.5-Embed-1B-mlx-8bit) — Embed-8B [8-bit](https://huggingface.co/netprojects/Nemotron-3.5-Embed-8B-mlx-8bit) · [4-bit](https://huggingface.co/netprojects/Nemotron-3.5-Embed-8B-mlx-4bit) — Rerank-8B [8-bit](https://huggingface.co/netprojects/Nemotron-3.5-Rerank-8B-mlx-8bit) · [4-bit](https://huggingface.co/netprojects/Nemotron-3.5-Rerank-8B-mlx-4bit) · [GGUF](https://huggingface.co/netprojects/Nemotron-3.5-Rerank-8B-GGUF) | Retrieval: embeddings and reranking for text and page images (scores match NVIDIA's published ViDoRe v3 results) |
27
+ | **H2O-Lightning 4B** | [q8](https://huggingface.co/netprojects/h2o-lightning-4b-mlx-q8) · [q4](https://huggingface.co/netprojects/h2o-lightning-4b-mlx-q4) | Small calibrated decision model from H2O.ai (text and images, two-way reading, Box questions) |
28
+ | **SenseNova-Loop-9B-Preview** | [8-bit](https://huggingface.co/netprojects/SenseNova-Loop-9B-Preview-MLX-8bit) · [6-bit](https://huggingface.co/netprojects/SenseNova-Loop-9B-Preview-MLX-6bit) · [4-bit](https://huggingface.co/netprojects/SenseNova-Loop-9B-Preview-MLX-4bit) | Looped Qwen3.5-based LM from SenseNova (layers 4–27 run twice, two-stream mHC) |
29
+ | **Fastino GLiDE 27B** | [8-bit](https://huggingface.co/netprojects/GLiDE-mlx-8bit) · [4-bit](https://huggingface.co/netprojects/GLiDE-mlx-4bit) | Calibrated decision model (choice / yes-no / score), midloop, full-attention reading |
30
  | **AutoTrust JEV-35B** | [8-bit](https://huggingface.co/netprojects/AutoTrust-JEV-35B-mlx-8bit) · [4-bit mixed](https://huggingface.co/netprojects/AutoTrust-JEV-35B-mlx-4bit-mixed) · [3-bit mixed](https://huggingface.co/netprojects/AutoTrust-JEV-35B-mlx-3bit-mixed) | AutoTrust decision and classification model |
31
  | **AutoTrust-JEV-27B v1.1** | [8-bit](https://huggingface.co/netprojects/AutoTrust-JEV-27B-v1.1-mlx-8bit) · [4-bit](https://huggingface.co/netprojects/AutoTrust-JEV-27B-v1.1-mlx-4bit) | Decision model with bidirectional prompt reading (custom loader included) |
32
+ | **IBM Granite Switch 4.1** | 8B [8-bit](https://huggingface.co/netprojects/granite-switch-4.1-8b-mlx-8bit) · [4-bit](https://huggingface.co/netprojects/granite-switch-4.1-8b-mlx-4bit) — 30B [8-bit](https://huggingface.co/netprojects/granite-switch-4.1-30b-mlx-8bit) · [4-bit](https://huggingface.co/netprojects/granite-switch-4.1-30b-mlx-4bit) | IBM Granite with switchable chat, decision and guardian modes |
33
+ | **NVIDIA Cosmos3-Edge reasoner** | [8-bit](https://huggingface.co/netprojects/Cosmos3-Edge-mlx-8bit) · [6-bit](https://huggingface.co/netprojects/Cosmos3-Edge-mlx-6bit) | Vision-language reasoner from NVIDIA: text, image and video in, text out |
34
+ | **NVIDIA Cosmos3-Edge video generator** | [8-bit](https://huggingface.co/netprojects/Cosmos3-Edge-Omni-mlx-8bit) · [bf16](https://huggingface.co/netprojects/Cosmos3-Edge-Omni-mlx-bf16) | Video generator from NVIDIA: text or a first frame in, video out (text-to-video, image-to-video, text-to-image) |
35
+ | **Qwen Concept 27B** | [bf16](https://huggingface.co/netprojects/Qwen-Concept-27B-MLX-bf16) · [8-bit](https://huggingface.co/netprojects/Qwen-Concept-27B-MLX-8bit) · [6-bit](https://huggingface.co/netprojects/Qwen-Concept-27B-MLX-6bit) (recommended) · [4-bit](https://huggingface.co/netprojects/Qwen-Concept-27B-MLX-4bit) | Qwen3.8-27B edited by Concept Country to reach answers with less thinking; text and images |
36
+ | **Laplace-1-Atlas** | [bf16](https://huggingface.co/netprojects/Laplace-1-Atlas-MLX-bf16) · [8-bit](https://huggingface.co/netprojects/Laplace-1-Atlas-MLX-8bit) · [6-bit](https://huggingface.co/netprojects/Laplace-1-Atlas-MLX-6bit) (recommended) · [4-bit](https://huggingface.co/netprojects/Laplace-1-Atlas-MLX-4bit) · [GGUF](https://huggingface.co/netprojects/Laplace-1-Atlas-GGUF) | Juneo Lab's 27B decision model: scores candidates in one forward pass; JevBench 91.8% (6-bit) |
37
 
38
  ## How we build
39