Horizon RunMap - claims and scope
Horizon RunMap is an experimental MoE execution architecture by Lucas Ribeiro. The structured claims record is the authority for supported claims, explicit non-claims and experiment scope. This page provides human-readable framing without creating a second record.
Supported claims
- selective-compression — Selective INT4 compression of routed experts with a separate FP16 non-routed path. Status: implemented-in-recorded-paths. Source.
- compact-execution — Compact expert GPU execution with FP16 activations and recorded scalar/grouped executors. Status: implemented-in-recorded-paths. Source.
- partial-residency — Partial residency observations under the selected campaign contract. Status: measured. Experiment: residency39.
- full-residency — OLMoE c64 full residency observations with Grouped T3. Status: measured. Experiment: residency39.
- instruction-following — Complete IFEval instruction-following evaluation of retained Horizon responses. Status: measured. Experiment: ifeval541.
Every measured claim remains bound to its named experiment, edition, model revision, executor, residency, hardware and workload contract.
What Horizon does not claim
- fastest MoE runtime.
- SOTA throughput.
- runtime-to-runtime superiority.
- superiority over llama.cpp.
- general BF16 quality equivalence.
- validated frontier-scale deployment.
- independently reproduced runtime generation.
- world-first quantization, offloading or caching.
The absence of a broader experiment limits the conclusions available from the publication; it does not refute a narrower claim the project does make.
Experiment inventory
| Experiment | Population | Evidence / validation / publication / reproduction |
|---|---|---|
| technical-preview-residency39-v3 | 13 configurations; 39 workers; 468 measured responses | E3 / valid / public / not_attempted |
| olmoe-0924-fidelity-v2 | 120 paired prompts | E3 / valid / public / not_attempted |
| olmoe-0924-ifeval541-v1 | 541 prompts; 834 instructions | E1 / valid / public / not_attempted |
The residency and executor campaign, paired structural study and IFEval benchmark are separate experiments. Their populations, scorers and research questions must not be merged into one quality or performance conclusion.
Interpretation boundaries
OLMoE partial-residency results and full-residency results use different recorded executors, so their difference is not a residency-only ablation. The published verifiers check only their declared artifact or score boundaries and do not rerun the private inference runtime. Missing observations remain UNKNOWN; unavailable capabilities remain UNSUPPORTED.
Machine-readable claims · Architecture record · Evidence · Related work