ChengyiX/ternary-bonsai2-27b-eval

Ternary Bonsai 2 27B on Apple Silicon

Long-Context Probe, KV-Cache Scaling, and Negative Results for the 20 tok/s Target

Chengyi Xu and Nexora — Nexora Research

---

Executive Summary

PrismML's Ternary Bonsai 2 27B uses end-to-end ternary weights (${-1, 0, +1}$) mapped to a compact 2-bit quantization (PQ2_0, 7.2 GB on disk) while maintaining near-full baseline accuracy on standard coding and reasoning benchmarks.

This repository publishes our experimental evaluation of Bonsai 2 on an Apple M4 Max 64 GB host under native 262,144 token context constraints.

Key Findings & Negative Results

1. Context Admissibility: The server successfully loaded and prefilled a 260,013-token prompt, retrieving needle sentinels with 100% precision.
2. Decode Throughput Collapse at Depth:
   • F16 KV Cache: Full-depth decode dropped to 3.37 tok/s (far below the 20 tok/s operational requirement).
   • Symmetric Q4 KV Cache: Decode achieved 5.29 tok/s at 260k tokens.
   • Short Context Baseline: Short-context reasoning (xhigh) ran at 21.87 tok/s.
3. Conclusion: While Ternary Bonsai 2's PQ2_0 weights fit comfortably within memory, KV-cache memory bandwidth and attention costs on Apple Silicon Metal architectures prevent simultaneous delivery of 262k context and ≥ 20 tok/s decode without specialized model kernels or memory layout optimizations.

---

Repository Contents

ternary-bonsai2-27b-eval/
├── README.md              # Overview and summary of findings
├── CITATION.cff           # Citation metadata
├── LICENSE                # MIT License
├── MANIFEST.json          # Deployment specification and environment metadata
├── RESEARCH.md            # In-depth architectural analysis and literature review
├── bench/
│   ├── context_probe.py   # Deterministic long-context probe harness
│   ├── test_context_probe.py # Probe unit test suite
│   ├── results.jsonl      # Machine-readable per-trial execution log
│   ├── context-260k-f16.json # Raw benchmark result (F16 KV)
│   ├── context-260k-q4.json  # Raw benchmark result (Q4 KV)
│   ├── context-260k-f16-no-think.json # Raw benchmark result (No reasoning)
│   ├── xhigh.json         # Raw benchmark result (Short prompt, xhigh reasoning)
│   └── tool_smoke_payload.json # Tool calling smoke test payload
├── deploy/
│   ├── start-canary.example.sh # Server launch script template
│   └── local.prism.bonsai2-262k.plist.example # macOS launchd agent template
└── logs/
    ├── canary-f16.log     # Sanitized startup and timing log (F16 KV)
    └── canary-q4.log      # Sanitized startup and timing log (Q4 KV)

---

Reproducing the Context Probe

Run the unit tests:

python3 bench/test_context_probe.py

Run a probe against an active OpenAI-compatible endpoint:

python3 bench/context_probe.py \
  --endpoint http://127.0.0.1:18795/v1 \
  --tokens 260000 \
  --needle "CRITICAL_SENTINEL_ALPHA" \
  --output probe-results.json

---

Citation

@dataset{xu2026bonsai2eval,
  title  = {Ternary Bonsai 2 27B on Apple Silicon: Long-Context Probe, KV-Cache Scaling and Negative Results for 20 TPS Target},
  author = {Xu, Chengyi and {{Nexora}}},
  year   = {2026},
  month  = oct,
  url    = {https://github.com/chengyixu/ternary-bonsai2-27b-eval}
}

---

License

MIT License. Copyright (c) 2026 Chengyi Xu and Nexora. Model weights are property of their respective creators (PrismML) and are not distributed in this repository.