A comprehensive reference and side-by-side configuration guide for deploying the qwen3.8:27b-mlx model across both Pi Agent (~/.pi/agent/models.json) and OpenCode (~/.config/opencode/opencode.jsonc) agent harnesses.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Show hidden characters
| { | |
| "$schema": "https://opencode.ai/config.json", | |
| "model": "ollama/qwen2.5-coder:7b", | |
| "provider": { |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| { | |
| "providers": { | |
| "ollama": { | |
| "name": "Local Ollama", | |
| "baseUrl": "http://localhost:11434/v1", | |
| "apiKey": "ollama", | |
| "api": "openai-completions", | |
| "authHeader": false, | |
| "headers": {}, | |
| "models": [ |
ThiagoEvoa
/ scratch_qwen_reference.md
Last active
November 4, 2026 00:10
Pi Agent Model Configuration Reference: qwen3.8:27b-mlx
ThiagoEvoa
/ ollama_apple_silicon_guide.md
Last active
August 19, 2026 18:11
A complete, production-grade guide for configuring a dedicated headless Apple Silicon Mac (M3, M4, and M5 generations) using Ollama. This setup is 100% agent-agnostic, serving a high-speed OpenAI-compatible API (/v1) for any client, coding assistant, or agent framework.
ThiagoEvoa
/ llm_context_limits_reference.md
Last active
August 21, 2026 12:14
A comprehensive, hardware-agnostic reference guide for configuring Large Language Model (LLM) context windows, output limits, Key-Value (KV) cache memory footprints, compaction strategies, and agent token budgets across Pi Agent, OpenCode, Ollama, and llama.cpp/vLLM.
Every LLM request operates inside a finite token budget:
ThiagoEvoa
/ apple_silicon_vram_optimization_guide.md
Created
August 14, 2026 18:35
This guide details how to maximize Unified Memory (vRAM) allocation for local LLM inference on Apple Silicon Macs (M-series) running macOS, specifically tailored for Ollama and high-demand workloads like Qwen models (qwen3.6:27b-mxfp8 and qwen3.6:35b-a3b-mxfp8).
By default, macOS allocates roughly 75%–80% of total unified memory to the GPU/vRAM, reserving the remainder for the system and desktop applications. When running large models with extended context windows (KV cache), manual overrides grant additional headroom to prevent out-of-memory errors or swapping.
ThiagoEvoa
/ local-model-configuration.md
Last active
September 26, 2026 15:55
models.json Configuration Guide
A unified, production-grade guide for configuring local and private LLM providers across both Pi Agent (~/.pi/agent/models.json) and OpenCode (~/.config/opencode/opencode.jsonc), structured across hardware RAM tiers (16 GB to 256 GB).
Both Pi Agent and OpenCode support running local models (via Ollama, MLX, llama.cpp, vLLM, or LM Studio) with OpenAI-compatible endpoints.
ThiagoEvoa
/ calc_ctx.py
Last active
August 15, 2026 18:45
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| #!/usr/bin/env python3 | |
| """ | |
| Ollama & LLM Safe Context Length Calculator | |
| Calculates recommended context length (num_ctx / max_ctx) based on VRAM, | |
| model architecture (GQA, layers, heads), weight size, and KV cache footprint. | |
| Includes dynamic Adaptive Safety Margin calculation based on model scale and live OS memory pressure. | |
| """ | |
| import argparse | |
| import json |
NewerOlder