Small models reason fine — they just don't know much. Parameters are the most expensive place to store facts: a fact costs ~100× more bytes in weights than in a compressed triple corpus. Ballast is that corpus: a versioned, CC0, rank-bucketed knowledge artifact you pair with any local model, quantized in levels like a GGUF.
Pull the corpus once, then every chat request gets grounded before your model sees it — fully offline after the pull:
uvx --from git+https://github.com/OpenBallast/ballast-cli openballast pull --level 3 # 265 MB uvx --from git+https://github.com/OpenBallast/ballast-cli openballast serve # point your client at http://localhost:11435/v1 instead of :11434
First live A/B on a 0.5B model: raw answer for Douglas Adams's birthplace was Dublin; through the proxy, Cambridge. Source: github.com/OpenBallast/ballast-cli (Apache-2.0).
Levels L0–L5 served live from a $0/month free-tier stack — demo-grade, no auth, no SLA. Point any MCP client at:
https://mcp.openballast.org/mcp
Or plain HTTP:
curl "https://mcp.openballast.org/lookup?question=Where+was+Douglas+Adams+born%3F&level=5" curl "https://mcp.openballast.org/evidence?id=Q42&level=3"
Canonical downloads live on Hugging Face — hive-partitioned parquet, 8 nested quantization levels (36 MB → 1.51 GB), CC0:
huggingface.co/datasets/OpenBallast/ballast-t0
Eval sets (50k linked probes): OpenBallast/ballast-evalsets · Thesis, methodology, numbers: github.com/OpenBallast/ballast
T0 = 197M cleaned Wikidata triples (CC0), entity-ranked by a composite of sitelinks and claim count, published as nested rank buckets. Truncation is the quality knob: drop deep buckets and evidence thins gracefully — measured rate–distortion, not vibes. A model-aware profile then selects tuned ballast: only what your model doesn't already know.