LLMRig

Know what your rig can run.

The open compatibility and performance intelligence layer for local AI.

LLMRig helps answer one practical question:

Which local LLM can this machine actually run well?

It inspects hardware, resolves logical models and runnable artifacts, identifies runtime paths, estimates compatibility, measures local execution, compares configurations, and preserves reproducible benchmark evidence. Curated setup uses Ollama; measured race execution supports Ollama, llama.cpp, and MLX-LM when compatible artifacts are already local.

Install

LLMRig is a CLI application, so the recommended installation method is pipx, which installs it in an isolated environment:

pipx install llmrig

Upgrade an existing installation:

pipx upgrade llmrig

Verify the installed CLI:

llmrig --version

Some system-managed Python installations, including common Homebrew Python setups, prevent global pip installs under PEP 668. If pipx is not available, use a virtual environment instead of modifying the system Python:

python3 -m venv .venv
source .venv/bin/activate  # macOS/Linux
python -m pip install llmrig

On Windows, activate the environment before installing with:

.venv\Scripts\Activate.ps1
python -m pip install llmrig

Then use LLMRig from anywhere:

llmrig doctor
llmrig can qwen3.8:27b-mlx
llmrig recommend
llmrig models --fit

Or run directly from source without installing:

python3 llmrig.py

LLMRig has no third-party Python runtime dependencies. The CLI uses only the Python standard library.

detect hardware
→ resolve model + artifact
→ identify runtime paths
→ estimate compatibility
→ execute locally
→ measure
→ compare
→ preserve evidence

LLMRig is currently Qwen-first. The architecture is intended to expand to additional model families, runtimes, GPUs, and platforms without changing the core workflow.

What LLMRig does

• Detects privacy-safe hardware facts across macOS, Windows, and Linux.
• Separates logical models from GGUF, MLX, Safetensors, and curated Ollama artifacts, with generic read-only Hugging Face resolution.
• Analyzes compatibility with explicit confidence, evidence provenance, practical context, and unknown handling.
• Detects Ollama, llama.cpp, and MLX-LM runtime capabilities while keeping installation and current availability separate from LLMRig adapter support.
• Recommends and sets up only curated, verified model identifiers through Ollama.
• Measures local race generation, prompt evaluation, and normalized inference latency through Ollama, llama.cpp, and MLX-LM adapters. The full bench workflow, residency measurement, and correctness smoke test remain Ollama-specific.
• Races at least two unique executable configurations with metric-specific, non-composite results.
• Exports privacy-safe benchmark passports containing raw per-run evidence and deterministic identities.
• Verifies passport structure, integrity, aggregates, and privacy entirely offline without inference.

Current scope

LLMRig remains Qwen-first. Its curated catalog supports practical recommendations and setup, while generic Hugging Face resolution inspects repository metadata without downloading model weights. Automatic installation remains limited to manually verified curated identifiers.

Ollama remains LLMRig's only setup/install backend. Race execution and benchmarking also support locally installed llama.cpp and MLX-LM runtimes. Those native adapters require explicit paths to already-local GGUF files or MLX model directories; LLMRig does not download native artifacts during a race.

The project name is intentionally broader than Qwen because the long-term direction is to support additional model families and runtimes without changing the user experience:

hardware → model + artifact → runtime paths → compatibility → measurement → evidence

If you want to add support for another model family, runtime, GPU vendor, or operating system, see CONTRIBUTING.md.

Requirements

• Python 3.9+
• macOS, Windows, or Linux
• Ollama for automatic model setup and the full benchmark workflow
• Optional llama.cpp or MLX-LM installations for native race execution
• Internet access for live discovery and model downloads
• No third-party Python runtime dependencies

Quick start

Run the interactive wizard:

llmrig

The wizard inspects the machine, recommends a supported model, pulls it if necessary, benchmarks it, and prints the local chat/API details.

Commands

Inspect the machine

llmrig doctor

Machine-readable output:

llmrig doctor --json

Check whether a curated model can run

llmrig can qwen3.8:27b-mlx

Machine-readable compatibility, confidence, configuration, and evidence:

llmrig can qwen3.8:27b-mlx --json

Curated identifiers receive the existing practical compatibility analysis. An owner/repository Hugging Face ID is resolved through read-only metadata and file listings, without downloading weights. LLMRig recognizes evidenced GGUF, MLX, and Safetensors artifacts. For generic GGUF and MLX artifacts it reports matching llama.cpp or MLX-LM capability candidates and whether those runtimes are locally available; Safetensors alone does not imply a runnable path. Runtime installation, current availability, format support, and total machine compatibility remain separate. Even a detected candidate is not a claim that inference was tested. Memory results are conservative planning estimates, and unmeasured runtime overhead, practical context, and performance remain explicitly unknown.

llmrig can also behaves as a three-state Unix predicate in both human and JSON modes: exit 0 means the model can run, exit 1 means it cannot run, and exit 2 means compatibility is unknown or the identifier cannot be analyzed.

Race locally executable configurations

llmrig race qwen3.8:27b-mlx
llmrig race qwen3.8:27b-mlx --json

Add an already-local MLX-LM artifact explicitly:

llmrig race <model> \
  --local-artifact mlx-lm=/path/to/local-mlx-model

Or add a local GGUF for llama.cpp:

llmrig race <model> \
  --local-artifact llama.cpp=/path/to/model.gguf

When the equivalent Ollama build is already installed, both native artifacts can be added for a three-way race:

llmrig race <model> \
  --local-artifact mlx-lm=/path/to/local-mlx-model \
  --local-artifact llama.cpp=/path/to/model.gguf

race measures only configurations that are already local, currently available, and backed by an LLMRig execution/benchmark adapter. It never installs runtimes or downloads artifacts. At least two executable configurations are required; otherwise the command reports the eligible and blocked alternatives without running a benchmark. At most one explicit artifact may be supplied for each native runtime. A GGUF target must be a non-empty local file; an MLX-LM target must contain an immediate config.json and at least one immediate non-empty model*.safetensors weights file. These are structural checks, not proof that the runtime can load the artifact. The local-artifact association is user-supplied evidence, not independent proof that differently packaged artifacts contain identical model weights or provide identical quality. LLMRig passes only explicit local paths to native runtimes and never invokes remote model identifiers for native race execution.

Exit 0 means at least two competitors were measured successfully. Exit 1 means an attempted execution failed and invalidated the comparison. Exit 2 means the race is unavailable or the model could not be resolved. Winners are reported separately for measured generation throughput, prompt-evaluation throughput, and normalized inference latency; there is no composite score or model-quality claim. Results within 5% are treated as inconclusive, and at least two timed runs per competitor are required for a winner.

Explain a measured decision

llmrig choose <model>
llmrig choose <model> --objective generation
llmrig choose <model> --objective prompt --json
llmrig choose <model> --objective latency
llmrig choose <model> --objective balanced

choose executes the same local comparison as race, then explains a decision for one explicit objective. Generation, prompt, and latency decisions reuse the existing metric-specific race results, including the two-sample requirement and 5% noise threshold. The default balanced objective is not a weighted score: it recommends a configuration only when exactly one configuration remains on the measured-performance Pareto frontier. Multiple frontier members make the decision genuinely inconclusive. Race warnings about quantization, artifact equivalence, formats, tokenization, early EOS, and user-supplied associations remain visible as caveats.

Exit 0 means a defensible recommendation was produced, exit 1 means race execution failed, and exit 2 means the decision is unavailable or inconclusive. choose never writes benchmark passports.

Optimize measured performance

llmrig optimize <model>
llmrig optimize <model> --json

optimize exposes an unranked, noise-aware Pareto frontier over measurements from the same local race path. The current dimensions are generation throughput and prompt- evaluation throughput (maximized), plus normalized inference latency (minimized). This is a measured-performance frontier, not a universal model-quality frontier. Memory is not yet a cross-runtime Pareto dimension, context is held constant by the race workload, and LLMRig does not infer quality, accuracy, or reasoning from speed. LLMRig does not create a universal runtime/model score.

A dimension participates only when every successful competitor has at least two valid measured samples and a positive finite value for it. Missing, invalid, or non-finite values are never treated as zero or as bad; the dimension is omitted globally and reported. Fewer than two remaining dimensions makes optimization inconclusive. One configuration dominates another only when it is not materially worse on every active dimension and is materially better on at least one, with differences within 5% treated as effectively tied. There is no composite score and frontier members are not ranked. Measured evidence takes precedence, and multiple frontier members mean the performance tradeoff remains unresolved.

Exit 0 means a valid frontier was derived from a completed race, exit 1 means race execution failed, and exit 2 means optimization is unavailable or inconclusive. optimize never writes benchmark passports.

In short: race measures, choose explains a decision for an explicit objective, and optimize exposes the measured-performance Pareto frontier. All three accept the same local race inputs; none downloads models or installs runtimes.

List models

Show curated local-ready models plus the newest live Qwen LLM/multimodal candidates:

llmrig models --fit

Force live refresh:

llmrig models --refresh --fit

Show the full Qwen Hugging Face organization catalog, including non-LLM artifacts:

llmrig models --all --fit

Use only the built-in curated snapshot:

llmrig models --offline --fit

Get a recommendation

llmrig recommend

Official models only:

llmrig recommend --category official

Community reduced-refusal models only:

llmrig recommend --category unrestricted

Prioritize quality:

llmrig recommend \
  --category official \
  --preference quality

For CLI convenience, unrestricted, uncensored, and reduced-refusal map to the community reduced-refusal category. restricted is accepted as an alias for the official category. LLMRig uses official and reduced-refusal in its output because those labels are more precise.

Set up a model

llmrig setup --category official

Or choose an exact curated model:

llmrig setup \
  --model qwen3.8:27b-mlx \
  --context 32768

If a known alias of the selected curated build is already installed, LLMRig reuses it when possible.

Benchmark installed models

One model:

llmrig bench \
  --model qwen3.8:27b-mlx \
  --context 32768 \
  --runs 2 \
  --passport benchmark.passport.json

All installed supported Qwen models:

llmrig bench \
  --all-installed \
  --context 32768 \
  --runs 2

LLMRig deduplicates installed aliases that resolve to the same Ollama model ID.

Benchmark passports

A benchmark passport is a versioned, privacy-safe JSON record of one measured execution configuration. It records the public model/build identifier, runtime, safe hardware summary, applied workload, individual timed samples, reproducible aggregates, and evidence provenance. Use bench --passport FILE for a single model, or race --passport-dir DIR to export each successfully measured race competitor.

Validate a passport locally without network access or inference:

llmrig passport verify benchmark.passport.json

Verification checks the schema, SHA-256 identity and configuration fingerprints, raw-sample aggregates, impossible states, and privacy constraints. It establishes only that the document is internally consistent according to LLMRig's schema. The hashes are identifiers and integrity checks, not signatures, independent proof, or benchmark certification.

passport_id is the SHA-256 hash of canonical passport content with passport_id itself excluded. It identifies the exact record, including its timestamp and measurements. configuration_fingerprint hashes the logical model, public artifact/build identifier, artifact digest when genuinely available, format, quantization, runtime and version, execution adapter, privacy-safe hardware facts, exact workload and generation settings, and benchmark method version. It excludes the passport ID, timestamp, measurements, aggregates, run-only warnings, and output path. The LLMRig tool version is record metadata; the benchmark method version is the compatibility boundary and must change when the procedure changes.

Two passports are exact when their configuration fingerprints match, even though their passport IDs and measured results may differ. comparable_with_warnings means the logical model and workload match but artifact, format, quantization, runtime, runtime version, or hardware differs. not_comparable means the logical model or workload—including context—differs. These classifications do not rank results. For user-supplied native targets, the public artifact identifier is path-independent and intentionally does not attest content identity. Matching native configuration fingerprints therefore do not prove that two user-supplied files contain the same weights.

Race passports are exported only when the overall race completes successfully. If any intended competitor fails, the race remains failed and --passport-dir writes no standalone competitor passports that could hide the incomplete comparison.

Passport aggregates are derived only from the recorded timed-run samples. Throughput means are rounded to two decimal places, wall-latency means to four decimal places, using Python's deterministic round behavior; generated tokens are the sum of the runtime-reported per-run counts. Warmups are recorded as policy metadata and never enter samples or aggregates.

Run project checks

Offline:

llmrig check

Include live Hugging Face discovery:

llmrig check --online

Model discovery and safety

LLMRig deliberately separates discovery from automatic installation.

The curated catalog contains local model identifiers that LLMRig may pull automatically. Live discovery queries the official Qwen organization on Hugging Face so new releases can appear without requiring an immediate LLMRig release.

A newly discovered repository is discovery only. LLMRig does not infer package size, hardware fit, or Ollama compatibility from a repository name. A model becomes eligible for one-command setup only after its identifier, package size, context capability, backend support, and provenance are verified and added to the curated catalog.

Community reduced-refusal discovery is best-effort and is not an authoritative registry. Third-party models should be reviewed before use.

Official vs reduced-refusal models

Official entries are upstream Qwen models distributed through the selected local backend.

Community reduced-refusal entries are third-party derivatives that modify model behavior to reduce refusals. Their authors may use terms such as uncensored, unrestricted, or abliterated.

Reduced refusal behavior does not imply better reasoning, accuracy, safety, or reliability. Review model provenance, licensing, and benchmark results before relying on a derivative for important work.

Hardware fit and context

LLMRig deliberately leaves headroom for the operating system, inference runtime, KV cache, other applications, and GPU/runtime buffers.

For Apple Silicon, CPU and GPU share unified memory. For discrete GPUs, LLMRig favors configurations likely to stay mostly or fully on the accelerator when VRAM can be detected.

The model's advertised maximum context is not automatically used as the default. Longer context allocations consume more memory, so LLMRig starts conservatively and increases context only when there is comfortable headroom.

Hardware-fit results are estimates. The benchmark on the user's actual machine is the final check.

Benchmarking

LLMRig unloads currently resident Ollama models before a benchmark and unloads the target model afterward. This reduces cross-model memory contamination and makes comparisons more reproducible.

Each benchmark records:

• generation tokens/second
• prompt-evaluation tokens/second
• load and total duration
• Ollama-reported context length
• accelerator residency when available
• RAM/swap snapshots when available
• three lightweight deterministic correctness smoke tests

Reports are written under benchmarks/ as JSON and Markdown. That directory is ignored by Git by default so local benchmark data is not published accidentally. Review any benchmark before sharing it.

These are local performance/configuration checks, not academic model-quality benchmarks. Memory values are snapshots rather than peak-memory measurements.

Contributing

LLMRig is open source and contributions are welcome.

Good first contribution areas include:

• additional model families
• new Ollama/local-backend model mappings
• AMD, Intel, and NVIDIA GPU detection improvements
• Windows and Linux hardware testing
• reproducible benchmark improvements
• new hardware profiles and recommendation rules
• documentation and usability improvements
• tests for new operating systems and model variants

Please read CONTRIBUTING.md before opening a pull request. Bug reports, feature ideas, model-support requests, and benchmark improvements are all welcome through GitHub issues.

Development

Run the full local validation set before opening a pull request:

python3 -m py_compile llmrig.py
python3 -m unittest discover -s tests -v
python3 llmrig.py check
python3 llmrig.py models --offline --fit

With internet access:

python3 llmrig.py check --online

GitHub Actions also runs compile, unit-test, and sanity-check jobs on Linux, macOS, and Windows.

Repository layout

llmrig/
├── .github/
│   ├── ISSUE_TEMPLATE/
│   ├── workflows/
│   │   ├── ci.yml
│   │   ├── publish-to-pypi.yml
│   │   └── refresh-profile-on-release.yml
│   └── PULL_REQUEST_TEMPLATE.md
├── assets/
│   └── llmrig-terminal.svg
├── tests/
│   ├── __init__.py
│   └── test_llmrig.py
├── .gitignore
├── CHANGELOG.md
├── CODE_OF_CONDUCT.md
├── CONTRIBUTING.md
├── LICENSE
├── pyproject.toml
├── README.md
├── SECURITY.md
└── llmrig.py

Roadmap

The roadmap is intentionally community-driven. Credible next directions include additional model families; llama.cpp, MLX-LM, and other execution adapters; richer GPU and runtime support; explainable comparison and decision intelligence; Pareto-style configuration optimization; and standardized privacy-safe benchmark sharing.

The rule for new functionality is simple: be useful, be reproducible, and do not turn unverified discovery metadata into an automatic install decision.

License

LLMRig is released under the MIT License.

Primary references

• Qwen official Hugging Face organization: https://huggingface.co/Qwen
• Qwen3.8 official repository: https://github.com/QwenLM/Qwen3.8
• Hugging Face Hub API: https://huggingface.co/docs/huggingface_hub/package_reference/hf_api
• Ollama documentation: https://docs.ollama.com/
• Ollama generate API: https://docs.ollama.com/api/generate
• Ollama running-model API: https://docs.ollama.com/api/ps
• Qwen3.8 Ollama tags: https://ollama.com/library/qwen3.8/tags