ornith

High-performance Lemonade alternative for AMD Strix Halo & Radeon — latest MoE models (Ling-3.0-Flash, Qwen 3.8 Flash Next, Ornith 1.5), bleeding-edge RDNA 3.5 Wave64/ROCmFP4 kernels, and silicon-tuned profiles.

A standalone local-only Python MCP server that gives any IDE persistent, workspace-isolated memory. works with any IDE supporting MCP servers

A Python CLI coding agent powered by remote Ollama, with workspace inspection and command-execution tools over Tailscale.

Ornith 1.5 Dropped: This 9B AI Coder BEATS Qwen 3.8? - Offline AI code generation scaffold using Ollama GGUF models.

Open models extended to 1M context with YaRN and certified needle by needle: Ornith, Gemma 4 uncensored, Qwen3.6 uncensored. MTP speculative decoding grafts, vision, full test harness.

A fully local OpenCode + llama.cpp + Ornith 1.5 coding-agent setup tuned for 128K context on a tight 6 GB VRAM budget.

Ornith-1.5-35B-A3B-NVFP4 on one DGX Spark with official vLLM

Benchmarking Edge-Based Large Models with Fewer Than 40 Billion Parameters. Includes updated models from vendors such as qwen, minicpm, Ornith, Spark, gemma, Ling, LFM, and RWKV.

Ornith-1.5-9B-MTP + KVMem on 8GB - KV-in-RAM long context 128K@42.7 tok/s, 210K@36.6, production config and limits

Pac-Man style HTML5 maze game generated by Ornith-1.0-9B on Jetson 8GB. 36/37 features, single-prompt generation.

Ornith-1.5-35B-A3B on RTX 4060 8GB - measured depth curves (PP/TG to 118K) + turbo4 KV production config + tool-loop fix

Ornith-1.5-122B (Qwen3.5-MoE 122B-A10B, vision) on one M5 Ultra: the publisher's own 4-bit MLX weights as published, TensorFold native Metal (PR #1125): 94-158 tok/s one stream with DFlash2 drafts, 2,480 tok/s prefill at 8.8k, 262k context, images, tools, structured output

Research & experiments running open-source LLMs locally on an 8GB VRAM laptop using Ollama/LM Studio with LangChain integration and performance benchmarks

Privacy-first guide and helper scripts for running Pi with local llama.cpp models on Apple Silicon.