Alia v0.1.0

First public release of Alia — a fully local coding agent for the GPU you already own.

alia detects your NVIDIA GPU, VRAM and RAM, picks the best measured hardware profile, starts llama-server on loopback and opens a branded Qwen Code terminal (or alia web for the browser). No API key, no cloud.

Measured on an RTX 3070 8 GB · Ryzen 7 5800X · 16 GB RAM (launch configuration)

Profile	Model	4K	32K	60K
rtx8gb-16gb-qwen36-35b (default, 64K context)	Qwen3.6 35B-A3B UD-IQ3_XXS + MTP	46.5 tok/s	35.2 tok/s	38.2 tok/s*
rtx8gb-qwen35-9b-fast (32K context)	Qwen3.5 9B Q4_K_M	66.1 tok/s	57.7 tok/s	—

* 60K from an earlier run with the same flags except --jinja. Long-context speed of the 35B profile depends on free RAM: under memory pressure it fell to 24.6 tok/s at 32K and 9.7 at 60K. 32 GB RAM is recommended for long contexts. Raw data: bench/results/.

Highlights

• Hardware-aware profiles with provenance for every number
• Safe server lifecycle: loopback only, PID + start-time identity, start lock, never kills foreign processes
• Expert offload + MTP speculative decoding for a 35B MoE on an 8 GB card
• English and German, following the system language
• Benchmark harness and raw JSONL results

Requirements

Linux, NVIDIA GPU, Python 3.12+, Qwen Code 0.25.1, llama.cpp ≥ a0e6dc7 built with CUDA. See the README.

Reviewed by CodeRabbit, the Codex connector and GitHub Copilot over four rounds; 47 unit tests, CodeQL clean.