metadata
language: en
license: mit
tags:
- benchmark
- long-context
- apple-silicon
- moe
- llama.cpp
- speculative-decoding
- agentic
size_categories:
- n<1K
ornith-256k-bench: Raw measurement records for LocalMoE (Ornith-1.5-35B-A3B on M4 Max)
Raw JSON-lines/JSON records backing the LocalMoE study: serving Ornith-1.5 35B-A3B (3B active). The original Q4_K_M experiment and the later Q6_K, reasoning-on comparison are both retained. The Q4_K_M long-context proof records a 260,013-prompt-token request on an Apple M4 Max (64 GB); the Q6_K record is a separate controlled short-workload comparison against a dense 27B + DFlash2 baseline.
Files (data/)
| File | Contents |
|---|---|
head_to_head_streaming_2026-08-26.jsonl |
Identical-prompt streaming suite, dense vs MoE, 5 categories × 3 runs |
head_to_head_streaming_q4vq6_2026-08-27.jsonl |
Q4_K_M versus pinned Q6_K comparison; Q6 labels are corrected to A3B |
head_to_head_pi_2026-08-26.jsonl |
pi agent tasks, both models, wall-clock + verification flags |
context_256k_proof.jsonl |
Near-limit sentinel proof (260,013 prompt tokens) |
raw_streaming.jsonl |
Full cold streaming suite, per-run records |
raw_streaming_cached_prefix.jsonl |
Prefix-cache reuse experiment |
raw_streaming_summary.json |
Cold-suite summary table |
qa_check_50k.json |
50k-context retrieval QA results |
agent_turns_24k.jsonl / .json |
Multi-turn prefix-reuse protocol records |
speed_research_summary.json |
Speculation-candidate evaluation matrix (negative results) |
pi_agent_suite/ |
Original six-task pi agent suite outputs |
Provenance
Single machine: MacBook Pro M4 Max 16-core, 64 GB unified memory, macOS 27.0. The reproducibility manifest pins the public Q6_K file, its SHA-256, llama.cpp revision, public paper/results checksums, and this dataset revision: https://github.com/minervacap2022/ornith-256k/tree/main/reproducibility
License
MIT. Model weights subject to upstream ornith.ai terms.