Skip to content

Arc Llama 1.0.0

Choose a tag to compare

@offbyonebit offbyonebit released this 24 Oct 01:09
· 69 commits to main since this release
3fc4ae1

Arc Llama 1.0 is the stable single-GPU release of the local LLM server for Intel Arc: GPU detection, runtime setup,
model management, and an OpenAI-compatible API. Validated natively on Linux and Windows 11 (Arc Pro B60).

pip install arc-llama
arc-llama run unsloth/Qwen3-8B-GGUF:Q4_K_M

Highlights

  • Redesigned web UI: Chat, Work, Models, Discover and System in one sidebar, with light/dark themes and mobile
    navigation.
  • Work (experimental): local project conversations with the built-in agent, OpenCode or Hermes Agent, run on
    isolated project copies with diff review, apply and undo, per-command approvals, acceptance checks and retries.
    Enable it with ARC_LLAMA_EXPERIMENTAL_AGENT=1 arc-llama serve.
  • Chat that lasts: reversible compaction, context budgets from the loaded model's own tokenizer, and answers that
    survive Stop, errors and dropped streams.
  • Per-model runtimes: a different llama-server executable and environment for each model.
  • Safer upgrades and rollbacks: settings written by newer versions are kept, and upgrades from 0.8 and 0.9 are
    tested against configs written by those releases.
  • SYCL without a oneAPI shell: runtimes and capability checks now source oneAPI themselves when its libraries are
    not on the loader path.
  • Windows: portability fixes and a recorded native Windows 11 validation run.

The full list is in the changelog.

Upgrading

Configs from 0.8 and 0.9 load unchanged. Before rolling back to an older version, see
Upgrading and rolling back.

Validation

  • CI on Linux and Windows, Python 3.10, 3.12 and 3.14; 1670 tests and an 88-case browser suite.
  • Native runs: Linux,
    Windows 11.
  • Clean install from the release wheel with an empty configuration: GPU detection, SYCL runtime, model registration
    and inference on an Arc Pro B60.

Checksums (SHA-256)

f2b11d418b09a9f8bf65e2aaf51f0528cbf78b478943e0b4f8458ab8751f42d2  arc_llama-1.0.0-py3-none-any.whl
80df1890721d539950bdb8c368d712a20262597d8a84b2c3a16238428d2947b3  arc_llama-1.0.0.tar.gz