Research paper · August 2026
Range Before
Representation
Behavior-gated two-bit quantization of a 35B mixture-of-experts model into a released 9.96 GB GGUF.
The MoE stayed useful because the range was chosen before the representation fight.
The controlled study held storage constant and changed how each group chose its quantization interval. Endpoint min/max scaling retained 77.1% of teacher-correct behaviors. Per-group MSE clip search retained 95.8% at the same 2.40 bpw simulated budget.
The exact release is separate from the simulation: a stock IQ2_XXS plus imatrix GGUF at 2.30 bpw. On the artifact gate, the bf16 teacher answered 118 items correctly and BTL-4 Compact reproduced 111 of them, for 94.1% conditional behavioral retention.
The 94.1% figure is conditional retention on a 120-item first-party behavior gate covering factual recall, grounded extraction, and false-premise rejection. It is not a general intelligence score, a coding result, or a claim that every capability retained 94.1%.
A falsification of the previous recipe, not a new codec.
BBP made the case for behavior-first promotion. RBR keeps that discipline but changes what the ablation teaches: for this routed model, expert tensors and range selection were the story.
- 01
Hold budget constant
Compare quantization decisions at the same nominal bit width, group size, and layer specification before assigning causality.
- 02
Choose the usable range
Endpoint min/max scaling wasted levels on outliers; per-group clip selection preserved much more measured behavior at the same storage.
- 03
Respect the level cliff
In this checkpoint, four expert-weight levels stayed close to the teacher while ternary and binary expert routes lost behavior sharply.
- 04
Measure where damage lives
BTL-3-style head and embedding protection did not transfer; the routed expert tensors dominated the measured quantization damage.
- 05
Ship the exact artifact
The release uses upstream IQ2_XXS GGUF plus an importance matrix, then gets judged as the physical file users actually run.
Standard format
A single GGUF using upstream llama.cpp IQ2_XXS weight types and an importance matrix, without a private packed representation.
Routed scale
A 35.1B-parameter MoE with roughly 2.1B active parameters per token, compressed into a 9.96 GB decimal file.
Expert sensitivity
Protecting output head and embedding did not rescue the model; measured damage was dominated by expert tensors.
Behavior gate
Retention was measured on the exact release artifact against teacher-correct, completed answers rather than tensor error alone.
Read the paper
The finished report is available as a PDF.
The paper records the controlled fake-quantization study, the shipped GGUF result, category-level retention, limitations, and the reason BRQ-style one-bit claims remain separate from this release.