FluxEM

Deterministic tools for exact tasks that small models still miss.

Status

FluxEM is now tools-first.

• Supported product surface: fluxem-tools-pkg/ -> fluxem-tools
• Supported release scope: a validated core of 35 tools across 9 domains
• Experimental or archived: fluxem/, old RL/research harnesses, research docs, and older benchmark artifacts

Archive details live in docs/ARCHIVE_STATUS.md.

What ships

fluxem-tools is a deterministic tool catalog for LLM tool calling. The package registry currently contains 210 tools across 42 domains, but the supported release surface is intentionally narrower:

• arithmetic
• number_theory
• physics
• chemistry
• biology
• statistics
• finance
• temporal
• information_theory

That validated core is the current product story. The broader catalog stays in-repo as experimental surface until it is benchmarked and documented to the same standard.

Repository layout

• fluxem-tools-pkg/ - supported package, release gate, catalog manifest, validation scripts
• examples/ - small-model and OpenAI-compatible integration examples
• benchmarks/ - current Qwen3.5-2B benchmark harness
• fluxem/ - archived research track
• experiments/ - archived or experimental evaluation/training work
• docs/ - package docs plus archived research pages

Quick start

pip install fluxem-tools

from fluxem_tools import call_tool

print(call_tool("arithmetic", expr="789123 * 456789"))
print(call_tool("physics_convert", value=88, from_unit="ft/s", to_unit="m/s"))
print(call_tool("finance_npv", rate=0.1, cashflows=[-100, 30, 40, 50, 60]))

SSD-backed Hugging Face cache

Benchmarks and local model downloads should use the SSD-backed Hugging Face root at /Volumes/VIXinSSD/huggingface.

export FLUXEM_HF_ROOT=/Volumes/VIXinSSD/huggingface
export HF_HOME=/Volumes/VIXinSSD/huggingface
export HF_HUB_CACHE=/Volumes/VIXinSSD/huggingface/hub
export HF_DATASETS_CACHE=/Volumes/VIXinSSD/huggingface/datasets
export TRANSFORMERS_CACHE=/Volumes/VIXinSSD/huggingface/hub

Local MLX quickstart (Apple Silicon)

Serve the validated model locally and run with retrieved tools:

# Start the MLX server
pip install mlx-lm
mlx_lm.server --model mlx-community/Qwen3.5-2B-OptiQ-4bit --port 8000

# Run a query
uv run python examples/qwen35_2b_openai_tools.py \
  "What is 789123 * 456789?" \
  --base-url http://127.0.0.1:8000/v1

Benchmark results

Qwen3.5-2B (4-bit, MLX) on 18 cases across 9 validated domains:

Condition	Accuracy	Avg Latency
Model only	38.9%	908 ms
Retrieved tools (max 8)	88.9%	911 ms
Full registry (210 tools)	72.2%	1302 ms

The retrieved-tools path is the recommended integration for models under 7B. Smaller tool surfaces produce better results on small models.

Entry points:

• examples/qwen35_2b_openai_tools.py - end-to-end example
• benchmarks/qwen35_2b_tools_benchmark.py - benchmark harness

Catalog and validation

The package now owns the release-facing catalog artifacts:

• fluxem-tools-pkg/catalog/manifest.json
• fluxem-tools-pkg/catalog/summary.md

To regenerate them:

cd fluxem-tools-pkg
python scripts/build_catalog_artifacts.py --strict-core

MCP server

pip install "fluxem-tools[mcp]"
fluxem-tools-mcp

The server exposes:

• fluxem_search
• fluxem_call

Research archive

The algebraic embedding work remains in the repository as archive material only. It is not the supported product story and it is not the release path for current FluxEM work.