hebrusInteractive and one-shot CLI
Open source / beta / built from source
Hebrus runs a deliberately small set of qualified model artifacts. Its AUTO planner chooses resident Metal or bounded SSD expert streaming against the machine and workload it can actually admit.
Qualified, not universal.
Hebrus supports exact artifacts and hardware profiles—not arbitrary GGUFs, model families or Macs. Unsupported combinations fail closed.Current support
Each row requires Apple Metal and the exact embedded ExpertMajor v2 store. Minimum memory is a qualified floor, not a promise that every context or workload fits.
| Model | Artifact | Minimum memory | Qualified AUTO path |
|---|---|---|---|
| Qwen3.6-35B-A3B Stable Affine4 | 20.81 GB | 16 GiB | Guarded SSD at 16 GiB through a 128K prompt plus 128 decode tokens. Higher-memory paths follow separate admission gates. |
| DeepSeek V4 Flash | 86.72 GB | 64 GiB | AUTO selects resident or SSD streaming after admission. |
| GLM 5.2 | 262.15 GB | 64 GiB | AUTO uses SSD streaming; resident requests are rejected. |
Runtime contract
The CLI, local server, alpha agent, benchmark and evaluation tools share the same runtime and machine-readable capability document.
Open the engine referencehebrusInteractive and one-shot CLI
hebrus-serverLocal HTTP server with model-specific protocol limits
hebrus-agentAlpha agent frontend
hebrus-benchBenchmark frontend
hebrus-evalEvaluation frontend
Model-free start
The repository documents Xcode Command Line Tools as the build prerequisite. These commands compile Hebrus and print its build and capability documents without downloading a model.
git clone https://github.com/andreaborio/hebrus.git
cd hebrus
git checkout v0.3.0
make -j
./hebrus --build-info
./hebrus --capabilities=jsonPrefer a desktop interface?