AI Models Overview

qwen3.6:27b-mlx

What it does

A multimodal model from Alibaba Cloud (27.8B parameters), optimized (4-bit quantized) for Apple's MLX framework. Features a dense architecture, hybrid attention mechanism, and a "thinking toggle" for chain-of-thought reasoning across conversational turns.

Best for

• Local Apple Silicon Workloads: Runs highly efficiently on modern Macs with 24GB+ unified memory.
• Agentic Coding: Excels at repository-level reasoning and frontend development workflows.
• Desktop Multimodal Tasks: Processes visual inputs natively without cloud APIs.

granite4.1:8b

What it does

Developed by IBM, this highly efficient dense 8B parameter model was trained on 15T tokens with a strict focus on data quality, governance, risk, and compliance (GRC). Supports 12 languages and boasts a 512K context window.

Best for

• Enterprise Applications: Apache 2.0 licensed and built on strict data clearance, making it highly safe for commercial deployment.
• RAG: Massive context window ideal for parsing deep internal company documents.
• Structured Outputs: Reliable for strict tool-calling and JSON extraction.

gemma4:26b

What it does

A frontier open multimodal model from Google DeepMind using a Mixture-of-Experts (MoE) architecture (25.2B total, activating ~3.8B per token). Processes text, images, and video, and incorporates an internal "Thinking Mode" for deep reasoning.

Best for

• High-Level Reasoning & Math: Punches well above its weight in logic and complex problem-solving.
• Consumer GPU Deployment: Runs fast on standard hardware due to low active parameter count.
• General AI Assistants: A highly versatile, locally hosted conversational experience.

gpt-oss:20b

What it does

OpenAI's open-weight MoE model (21B total, 3.6B active). Trained using reinforcement learning informed by proprietary o-series reasoning models. Features an adjustable "reasoning effort" parameter.

Best for

• Edge Devices & Local Hardware: Requires only ~16GB of memory to run efficiently.
• Advanced Tool Use: Optimized for deep research, Python execution, and complex multi-step tool calling.
• CoT Dominance: Delivers raw, academic-level mathematical and coding reasoning.

yi-coder:9b

What it does

A highly efficient open-source code generation model (8.83B parameters). Trained specifically for software development across 52 languages, featuring a massive 128K token context window.

Best for

• Repository-Level Understanding: Can digest entire codebases and dependency trees to hunt for bugs.
• Local Coding Assistance: Lightweight enough to provide fast, private code completion on consumer hardware.

LLaVA

What it does

Large Language and Vision Assistant. A foundational multimodal model that connects a visual encoder with an LLM backbone (like LLaMA), fine-tuned using multimodal instruction-following data.

Best for

• Visual Q&A: Excellent at answering open-ended conversational questions about image contents.
• OCR: Highly effective at reading and interpreting text inside images and diagrams.
• Custom Vision Apps: A powerful open-source foundation for building custom vision-language tools.

codegemma

What it does

Google's lightweight code models built on the Gemma architecture. Comes in 7B and 2B variants, offering native support for "fill-in-the-middle" (FIM) code operations.

Best for

• IDE Autocomplete: FIM allows it to seamlessly inject completions based on surrounding context.
• Boilerplate Reduction: Generates syntactically correct code in major languages to speed up development.
• Ultra-Fast Local Generation: The 2B variant acts as a highly responsive, ultra-fast local co-pilot.

qwen2.5vl:7b

What it does

Alibaba's advanced 7B parameter vision-language model. Fuses a custom Vision Transformer with a language decoder to process high-resolution images/videos dynamically, supporting a 256K+ context window.

Best for

• Spatial & Temporal Grounding: Generates exact bounding box coordinates and analyzes events across long videos.
• Complex Document Parsing: Perfect for handling high-resolution PDFs, charts, and structured data extraction.
• Multimodal Agents: Acts as the "brain" for agents navigating graphical GUIs or the visual web.

Whyimhere/ReaderLM-v2

What it does

• • ReaderLM‑v2 (from Jina AI) is a 1.5B-parameter language model specialized in converting raw HTML into clean Markdown or structured JSON.
• • It treats HTML→Markdown as a translation task rather than selective copy, so it preserves structure (headings, lists, tables, code fences, LaTeX, etc.) with high fidelity and less “degeneration” on long outputs.
• • It supports very long contexts (up to around 512K tokens combined input+output), and 29 languages, which makes it suitable for whole‑site or doc‑set processing rather than just snippets

Best for

• Automated HTML → Markdown conversion pipelines for blogs, docs, knowledge bases, or archives. ◦ Example: scrape article HTML, feed to ReaderLM‑v2, store clean Markdown in your CMS or a static‑site generator repo.
• HTML → JSON extraction where you define a schema (e.g., {title, author, date, sections, code_blocks}) and have the model output structured data directly, skipping an intermediate Markdown step.
• Large‑scale web content processing: main‑content extraction, stripping boilerplate, and building datasets/indices (e.g., for search or RAG) from heterogeneous HTML sources.

Whyimhere / Oxy‑1‑micro

What it does

• Oxy‑1‑micro is a fine‑tuned variant of Qwen2‑1.5B, targeted specifically at role‑play and interactive dialogue
• The finetuning focuses on generating dynamic, in‑character, context‑rich conversations and interactive storytelling, prioritizing style and persona over raw reasoning power
• Its small size makes it fast and cheap to run locally or on modest hardware while still producing relatively engaging, coherent dialog for its niche

Best for

• Character and NPC chat: in‑game companions, narrative bots, or chat characters where sustained persona and style matter more than complex logic.
• Interactive fiction and RP: collaborative storytelling, text adventures, or chat‑based role‑play experiences hosted on the web or in Discord‑style bots
• Lightweight, always‑on RP/chat agents on low‑resource servers or local machines, where you want fast response and don’t need full GPT‑class capabilities