model_architecture string | quantization_notes dict | perplexity_benchmarks dict | metal_gpu_throughput dict | 70B_ceiling dict | timestamp_utc timestamp[s] |
|---|---|---|---|---|---|
BitNet b1.58 Ternary Matrix | {
"bits_per_parameter_theory": 1.58,
"bits_per_parameter_theory_derivation": "log2(3) — Shannon entropy of one ternary weight {-1,0,+1}; the theoretical lower bound, NOT the in-memory layout.",
"packed_layout_bits_per_weight": 2,
"memory_reduction_vs_fp16": 8.1,
"memory_reduction_vs_fp16_derivation": "16 / 2 ... | {
"protocol": {
"model": "Qwen/Qwen2.5-0.5B (BitNet b1.58 continued-train, masked CE, deployed-forward QAT)",
"parameter_count": "0.5B",
"context_length": 256,
"tokenizer": "Qwen tokenizer (Qwen/Qwen2.5-0.5B)",
"note": "The three rows below are slices of the SAME training file (konstellation corpu... | {
"device": "Apple M3 Max",
"device_sustained_memory_bandwidth_gbps": 384.2,
"note_bandwidth": "Device peak sustained bandwidth (M3 Max memory system). This is a hardware ceiling, NOT a decode measurement.",
"measurement": {
"model": "Qwen/Qwen2.5-0.5B quantal-ternary, B8/256, single-stream decode",
"to... | {
"weights_per_token_gb": 17.28,
"derivation": "70e9 params x 1.975 bits/weight / 8 / 1e9",
"tok_sec_at_peak_bandwidth": 22.2,
"derivation_tok": "384.2 / 17.28",
"note": "This is a CEILING for what 70B would do IF it became memory-bound at peak bandwidth. The 0.5B run in this file (142.8 tok/s, 4.6% of peak) ... | 2026-08-11T01:11:05 |
kompress-ultra-bitnet-benchmarks
BitNet b1.58 ternary quantization metrics: memory reduction, perplexity, and Apple Silicon Metal throughput.
Files
bitnet_b158_compression_results.json— the metrics payload.
Denominators (read these first)
The JSON intentionally carries two compression numbers because they mean different things:
| Number | Meaning | Denominator |
|---|---|---|
| 8.1x | measured memory reduction vs fp16 | 2 bits/weight (packed ternary layout; 16/2 = 8.0x + packing overhead) — the honest headline |
| 1.58 | theoretical entropy of one ternary weight | log2(3) = 1.585 — the information-theoretic bound, NOT the shipped layout |
16 / 1.58 = 10.13x and 16 / 8.1 = 1.975 bits are both real, they just
answer different questions. The JSON labels each explicitly.
Protocol
- Model: Qwen/Qwen2.5-0.5B BitNet b1.58 continued-train (masked CE, deployed-forward QAT), exported as 168 ternary matrices.
- Context: 256 tokens · Val: 90-sample stratified held-out.
- Throughput — two measurements, not one:
- 384.2 GB/s is the M3 Max device peak sustained bandwidth (hardware ceiling; M4 Pro is a different memory system, NOT folded in).
- 142.8 tok/s is the 0.5B quantal model's actual single-stream decode. At this size that is 4.6% of peak (0.1234 GB weights/token × 142.8 = 17.6 GB/s) — the decode is overhead-bound, NOT memory-bound. A memory-bound ternary decode would need a model ~20× larger.
- 70B ceiling: 70e9 × 1.975 bits / 8 = 17.28 GB weights/token → ~22 tok/s at 384 GB/s. This is a hardware ceiling, not a scale-check of the 0.5B run — the 0.5B model is nowhere near it (4.6% of peak).
- Downloads last month
- 53