Alia
GitHub
v0.2.0 · open source · Apache-2.0

A local coding agent
for the GPU you already own.

Alia detects your GPU, picks a benchmarked model profile, starts llama.cpp and opens a full coding agent in your terminal. No API key. No cloud. Nothing leaves your machine.

Get started View on GitHub

Linux · NVIDIA · tested on an RTX 3070 with 8 GB VRAM

Three real prompts on an RTX 3070, no edits by hand: Alia writes a live GPU dashboard (one Python file, standard library only), polishes the layout, then finds and fixes its own bug with grep. Terminal parts are sped up as labelled; the polish round was slow because the server was swapping. The last clip is the result running in real time.

49.9 tok/sgeneration with a 35B MoE model at 4K context, exact launch configuration
8 GBVRAM is enough: experts stream from system RAM
10.9Ktokens first prompt (was 22.8K)
0API keys, accounts or cloud calls

Why Alia

Everything a coding agent needs.
Nothing that phones home.

Hardware-aware profiles

Alia reads VRAM and RAM, then starts the best measured profile for your machine. Pin another one with alia use, compare them with alia profiles.

35B on an 8 GB card

A mixture-of-experts model activates only 3B parameters per token. Attention stays on the GPU, experts are offloaded to RAM.

Speculative decoding

The model's own MTP head drafts two tokens ahead. 65 % are accepted: +14 % speed at 4K, +10 % at 32K.

A real agent

Built on Qwen Code: it edits files, runs commands, reads errors and fixes them. Resume anytime with alia -c. It asks before acting unless you choose --auto-edit or --yolo.

Terminal or browser

alia web opens the same agent as a web shell, bound to 127.0.0.1 only.

Evidence, not claims

Every benchmark number on this page links to raw JSONL; configuration gaps are identified explicitly.

Lean by design

17 unused tools denied, no background memory calls, no follow-up requests, skills in ~/.agents/skills not loaded. The first prompt shrank from 22.8K to 10.9K tokens. A one-word reply on a warm server takes ~3 s (was 7 s).

Browser and web search, opt-in

Optional via MCP, off by default: alia mcp enable browser (Playwright MCP, text snapshots, no arbitrary JavaScript tool) and alia mcp enable search (DuckDuckGo, no API key, rate-limited). alia mcp test checks both. In tests, the local 35B model read a JavaScript-rendered page in 26 s and found a package version with source in 24 s.

Benchmarks

Measured, not estimated.

Token generation speed as the context fills up, with each profile's exact launch configuration. RTX 3070 8 GB, Ryzen 7 5800X, 16 GB RAM, llama.cpp a0e6dc7.

4K context
49.9
66.2
32K context
39.1
57.8
60K context
37.4
51.4

Qwen3.6 35B-A3B

Code tasks
9/10
Tool calls
7/8
JSON with / without schema
3/3 · 3/3
Prompt speed 4K / 32K
1,034 / 1,451 tok/s
Download
14.07 GB

Qwen3.5 9B

Code tasks
8/10*
Tool calls
8/8*
JSON schema
—
Prompt speed 4K / 32K
2,964 / 2,695 tok/s
Download
5.7 GB

* Measured via Ollama with Qwen's own GGUF, not the unsloth GGUF and llama-server configuration this profile launches.

Agent tasks by profile

Share of hidden tests passed on real repository tasks. Longer bar = more tasks solved.

Ornith-1.5 35Bquality · opt-in16/16
248 s per run · 41.8 tok/s
Qwen3.6 35Bdefault13/16
104 s per run · 54.6 tok/s · 2.4× faster
Qwen3.6 35B · 128Klong context · opt-in3/3
3 of the 8 tasks · 68 s per run · 44.1 tok/s · compacts at ~98K instead of ~32.5K

Eight frozen repository tasks, two runs each, solved headless by Qwen Code and scored by hidden tests; time is the mean per run. Switch with alia use <profile>, back with alia use auto.

All speeds use exactly the flags and context size each profile launches with (65,536 tokens for both). Long-context speed needs free RAM: with a browser and other apps open, the 35B model streamed from the SSD and fell to 24.6 tok/s at 32K and 9.7 at 60K. 32 GB RAM is recommended for long contexts. Quality scores come from separate runs. Raw results and methodology →

In use

Your terminal. Or your browser.

Alia in the terminal: the agent writes primes.py, the first run fails because python is missing, it retries with python3 and succeeds

Showcase

Alia, explained by Alia.

We asked Alia to visualize its own architecture. Running locally on the RTX 3070, it read its own source and wrote the whole page: an interactive diagram, an animated request flow and the measured numbers. Every prompt is published.

Architecture page written by Alia: hero titled Alia, explained by Alia, and a layered diagram from terminal and browser through the Alia launcher to Qwen Code, llama-server and GPU/RAM
Open the page Read the prompts

How it is built

Standing on proven open source.

Alia does not reinvent the agent loop or the inference engine. It adds what was missing: hardware detection, measured profiles, a safe server lifecycle and one command to start it all.

  1. Aliahardware detection · profiles · server lifecycle · branding
  2. Qwen Codeagent loop, tools, terminal UI, web shell
  3. llama.cppCUDA inference, expert offload, MTP drafting
  4. Qwen3.6 · Unsloth GGUFopen-weight model, dynamic 3-bit quantization
  5. Your GPUNVIDIA · Linux

Install

Up and running in four steps.

  1. Agent and downloader

    npm install -g @qwen-code/qwen-code@0.25.1
python3 -m pip install -U "huggingface_hub[cli]"

  2. llama.cpp with CUDA

    git clone https://github.com/ggml-org/llama.cpp ~/src/llama.cpp
cd ~/src/llama.cpp && git checkout a0e6dc7
cmake -S . -B build -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build -j"$(nproc)" --target llama-server

  3. Alia

    git clone https://github.com/BEKO2210/alia
cd alia && ./install.sh && alia doctor

  4. Start & commands

    alia        # start agent
alia -c     # resume session
alia status # server PID, profile, uptime
alia mcp    # browser & web search

    The first start shows the exact model download (14.07 GB for the default profile) and asks before fetching anything. Check status with alia status, continue with alia -c, or manage tools with alia mcp.

Own your coding agent.

Free, open source, and it runs on the hardware you already have.

Star on GitHub Add your GPU profile