AI & ML interests

On-device and local LLMs on Apple Silicon (MLX), small calibrated decision models, fact-checking and content quality for long-running websites, and open-source tooling for small teams.

Recent Activity

Paris99  updated a collection about 7 hours ago
NetProjects GGUF builds
Paris99  updated a model about 7 hours ago
netprojects/Qwen3.8-27B-hermes-adapted-GGUF
Paris99  published a model about 7 hours ago
netprojects/Qwen3.8-27B-hermes-adapted-GGUF
View all activity

Organization Card

NetProjects

Open models, ready to run on Apple Silicon.

We convert and test open models for MLX (and GGUF for llama.cpp), so they run locally on a Mac with no cloud and no API keys. We check every build against the original model before it goes up, and we publish speed numbers from real hardware (MacBook Pro M5 Max, 128 GB).

Models

Family Builds What it's for
NVIDIA Nemotron Dory 1.0 4B bf16 · 8-bit · NVFP4 4-bit Looped reasoning LM from NVIDIA Nemotron Labs
Tencent DSpine drafter for Qwen3.8-27B drafter 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
NVIDIA Nemotron 3.5 Embed / Rerank Embed-1B bf16 · 8-bit — Embed-8B 8-bit · 4-bit — Rerank-8B 8-bit · 4-bit · GGUF Retrieval: embeddings and reranking for text and page images (scores match NVIDIA's published ViDoRe v3 results)
H2O-Lightning 4B q8 · q4 Small calibrated decision model from H2O.ai (text and images, two-way reading, Box questions)
SenseNova-Loop-9B bf16 · 8-bit · 6-bit (recommended) · 4-bit — Preview: 8-bit · 6-bit · 4-bit Looped Qwen3.5-based model from SenseNova (layers 4–27 run twice, two-stream mHC): text + image input, MTP decoding ~1.7× faster, 512K context
Fastino GLiDE 27B 8-bit · 4-bit Calibrated decision model (choice / yes-no / score), midloop, full-attention reading
AutoTrust JEV-35B 8-bit · 4-bit mixed · 3-bit mixed AutoTrust decision and classification model
AutoTrust-JEV-27B v1.1 8-bit · 4-bit Decision model with bidirectional prompt reading (custom loader included)
IBM Granite Switch 4.1 8B 8-bit · 4-bit — 30B 8-bit · 4-bit IBM Granite with switchable chat, decision and guardian modes
NVIDIA Cosmos3-Edge reasoner 8-bit · 6-bit Vision-language reasoner from NVIDIA: text, image and video in, text out
NVIDIA Cosmos3-Edge video generator 8-bit · bf16 Video generator from NVIDIA: text or a first frame in, video out (text-to-video, image-to-video, text-to-image)
Qwen Concept 27B bf16 · 8-bit · 6-bit (recommended) · 4-bit Qwen3.8-27B edited by Concept Country to reach answers with less thinking; text and images
Laplace-1-Atlas bf16 · 8-bit · 6-bit (recommended) · 4-bit · GGUF Juneo Lab's 27B decision model: scores candidates in one forward pass; JevBench 91.8% (6-bit)
Laplace-1-Vector bf16 · 8-bit · 6-bit (recommended) · 4-bit Juneo Lab's 4B decision model: scores candidates in one forward pass; JevBench 79.2% (6-bit)
SenseNova-6.8-Code2Video bf16 · 8-bit · 6-bit (recommended) · 4-bit SenseNova's 27B agentic code-to-video model: writes, renders and reviews video as code; vision tower kept
Mage-VL-Cowork bf16 · 8-bit · 6-bit (recommended) · 4-bit Microsoft's 4B hybrid computer-use agent (GUI + bash tools); vision tower kept in bf16
Qwen3.8-27B-pi-adapted 6-bit (recommended) · 8-bit · 4-bit · 3-bit · 2.5-bit · bf16 · MLX adapter · GGUF harness-hub's pi coding-agent LoRA merged into Qwen3.8-27B (Terminal-Bench 2.1 39.5% → 55.8% under pi, publisher's figures); thinking off, vision and MTP kept
Qwen3.8-27B-opencode-adapted 6-bit (recommended) · 8-bit · 4-bit · 3-bit · bf16 · MLX adapter · GGUF harness-hub's OpenCode coding-agent LoRA merged into Qwen3.8-27B (SWE-bench Pro hard-51 35.3% → 51.0% under OpenCode, publisher's figures); thinking off, vision and MTP kept
DaisyRose-27B 6-bit (recommended) · 8-bit · 4-bit · 3-bit · 2.5-bit · bf16 RadixArk's reinforcement-learned Qwen3.8-27B (Terminal Bench 4.1 5.6 → 12.0, publisher's figures); thinking model, vision and MTP kept
Qwen3.8-27B-hermes-adapted 6-bit (recommended) · 8-bit · 4-bit · 3-bit · 2.5-bit · bf16 · MLX adapter · GGUF harness-hub's Hermes Agent LoRA merged into Qwen3.8-27B (SWE-bench Pro hard-51 29.4% → 43.1% under Hermes Agent, publisher's figures); thinking off, vision and MTP kept
Arete-27B 6-bit (recommended) · 8-bit · 4-bit · 3-bit · 2.5-bit · bf16 Nous Research's fine-tune of Qwen3.8-27B for factual, direct answers on sensitive topics; vision kept and MTP; 2.5-bit build is text-only

How we build

  • Conversions use the standard mlx-lm / mlx-vlm tools, or llama.cpp for GGUF builds; where a model needs custom code, the loader is open source on GitHub.
  • Each card says which upstream revision it was built from, how it was quantized and how it was checked, with speed and memory measured on an M5 Max.
  • Licences follow the original model.

Code

All builds: MLX collection and GGUF collection.

Also

We also work on small calibrated decision models, fact-checking and content quality for long-running websites, and open-source tooling for small teams.

Spotted a problem or want a model converted? Open a discussion on any model page.

datasets 0

None public yet