aris-server 0.1.0 · on PyPI

Your coding agent, running on your Mac.

Aris Server runs open-weight models on Apple Silicon and serves them the way coding agents expect: the Anthropic Messages API, for Claude Code, Aris Code and anything else that speaks it. Aris 3.8-27B, our open coding-agent model, is built in.

  • Free and open source, Apache-2.0.
  • Private: the model and everything you send it stay on your Mac.
  • In Aris Code, one click: Settings → Providers → Aris Local.
install, download, serve
uv tool install aris-server
aris-server pull aris-3.8-27b
aris-server serve --model aris-3.8-27b
then, in Claude Code
ANTHROPIC_BASE_URL=http://127.0.0.1:8484 ANTHROPIC_AUTH_TOKEN=local claude

pull sizes every build of the model from Hugging Face before anything downloads, and fetches the one this Mac holds.

//The model

Aris 3.8-27B

An open coding-agent model. It learned from its own agent runs that passed their tests: write the code, run the tests, fix what fails, then stop. The 3.8 is Qwen's version number.

  1. Base model
    Qwen3.8-27B
    The Qwen team, Alibaba Cloud
  2. Fine-tune
    Arete-27B
    Nous Research
  3. Ours
    Aris 3.8-27B
    Trained on its own passing agent runs
27 GB
of weights at 8 bits
~125k
tokens of context on a 48 GB Mac
~8 tok/s
decode on an M4 Pro
Apache-2.0
open weights

Same accuracy, less second-guessing

76 reasoning items at high effort: 60 MATH-500 problems, and 16 questions that apply a spec's rules.

Arete-27BAris 3.8-27B
Correct71/7672/76
Self-corrections per 1,000 tokens2.952.57
Answers that hit the token cap00

The accuracy difference is within noise. "Self-corrections" counts "wait", "hmm" and "let me reconsider" while it thinks: about 13% fewer.

Agent tasks, graded by hidden tests

Claude Code drove the model through four libraries built from a spec. It never saw the tests.

TaskLanguageArete-27BAris 3.8-27B
JSON Pointer and JSON PatchTypeScript34/35 ¹35/35
npm semver range matchingRust50/5250/52
Expression evaluatorRust45/4545/45
.gitignore matchingPython64/64 ²64/64

Aris finished all four without help. ¹ Arete got a hint mid-run. ² Resumed by hand after the server ran out of memory. Each task was run once.

Long context holds up: 18 of 18 questions right about a 50k-token synthetic codebase, and 17 of 18 at 100k. It inherits Arete's value-alignment training and Qwen3.8's knowledge cutoff. The model card has the full method and every caveat.

Model card on Hugging Face →

Apache-2.0, like Arete-27B and Qwen3.8-27B. Built on the work of the Qwen team at Alibaba Cloud and Nous Research: thank you both. Training tasks from Exercism's Python and Rust tracks (MIT).

//In Aris Code

One click, no terminal.

Aris Code sets the whole thing up from its settings: it installs Aris Server, downloads Aris 3.8-27B with a progress bar you can cancel and resume, starts the login service and points its Aris Local provider at it. The knowledge graph tools come along, like on every other provider.

Aris Code · Settings
Providers / Aris Local
Local model

Installs Aris Server and downloads Aris 3.8-27B (about 29 GB) to run on this Mac. Needs Apple Silicon and 48 GB of memory.

Cancel
Downloading Aris 3.8-27B: 17.7 of 28.6 GB.
Install
Aris Server
Download
the build this Mac holds
Run
as a login service
//Why not Ollama?

One job, done properly.

Serve a local model to coding agents.

Native /v1/messages

Thinking, tool calls and text stream as real Anthropic events from the first byte, with pings while a long prompt is read, so Claude Code never times out on silence. A client that hangs up stops its generation.

Effort that works

Claude Code's effort setting reaches models with trained effort levels, and a think budget closes a reasoning model's thinking at a line break once it has thought enough.

The right build for each Mac

Give a model several builds and the server works out, before anything downloads, which one this Mac holds with room for a long conversation.

Long conversations stay warm

Prompt caches carry the system prompt and each conversation from turn to turn, and onto disk, so a restart picks up where it left off instead of re-reading 100k tokens.

Twice the context

An 8-bit KV cache that keeps batching: about twice the context in the same memory, with the same answers.

Runs as a service

It starts at login and comes back by itself after any exit, so the model is there when your agent needs it.

Built on Apple's MLX and mlx-lm. macOS on Apple Silicon only, for now. Aris 3.8-27B's 8-bit build needs 48 GB of memory, and a 4-bit build for 32 GB Macs is coming. Any MLX model works by name: aris-server serve --model mlx-community/Qwen3.5-9B-MLX-8bit.

Read the source.

The server, its design notes and its tests are on GitHub.