A ~0.5 GB ternary Qwen3-1.7B whose reasoning quality was restored by
QAT healing — the flagship result of a systematic 16-arm research
campaign. Weights are 2-bit ternary (grid identical to
prism-ml/Ternary-Bonsai-1.7B);
the healing moved ~0.1–0.2 % of grid codes via straight-through-estimator
training around the frozen ternary grid.
Variants
File
Focus
Size
Permea2.0-1.7B-P.Q2_0.gguf
Benchmark champion — best external scores
0.49 GB
Permea2.0-1.7B-N.Q2_0.gguf
Balanced — strongest instruction-holdout
0.49 GB
The core finding: doubt loops, and how to cut them
Naive ternary quantization does not destroy knowledge — it destroys
commitment. The quantized model reaches correct answers in its
reasoning trace, then cycles "But wait, let me check again…" until any
token budget is exhausted. We showed this failure mode is not fixable
by larger budgets or anti-repetition samplers — only by training:
the student mines its own truncated loop traces on training data,
each trace is cut at the first provably-correct commitment point, and
the repaired traces are distilled back. ~200 such repair pairs recover
most of the gap to the bf16 base.
Capability vs. size
Benchmarks (zero-shot, identical harness/items/budget for all rows)
Per-benchmark panels
Measured with our open evaluation harness: 550 items
(GSM8K 200 / MMLU 200 / ARC-Challenge 150, official test splits),
generation budget 3072 tokens, greedy, answer extraction with
boxed/statement fallbacks. Numbers are paired — every model saw the
identical items through the identical grader.
Model
Size
GSM8K
MMLU
ARC-C
Overall
Permea2.0-1.7B-P (this)
0.49 GB
71.5
46.5
69.3
61.8
Qwen3-0.6B (Q4_K_M)
0.40 GB
61.0
48.5
69.3
58.7
SmolLM2-360M (Q8)
0.39 GB
4.5
31.0
22.7
19.1
Ternary-Bonsai-1.7B (unhealed)
0.5 GB
68.5
46.5
64.0
59.3
Qwen3-1.7B bf16 (teacher/base)
3.4 GB
81.0
61.5
88.0
75.8
Qwen3-0.6B (Q8)
0.64 GB
69.5
48.0
68.7
61.5
Llama-3.2-1B-Instruct (Q8)
1.31 GB
33.5
29.5
35.3
32.5
Gemma-3-1B-it (Q8)
1.07 GB
3.0
33.5
45.3
25.6
Reasoning-capable models (ours, Qwen3-0.6B) were measured in thinking
mode; Llama-3.2-1B and Gemma-3-1B are non-reasoning models and were
measured with their standard chat templates — same items, budget and
grader for every row. Competitor GGUFs are pinned to fixed revisions.
Notes: published third-party numbers are not directly comparable
(few-shot prompts, different harnesses). The honest comparison is the
paired delta inside this table. Our harness reproduces published
Qwen3-1.7B anchors within ~1–5 points (GSM8K 81.0 vs 75.4 published,
MMLU 61.5 vs 62.6).
2.1× faster than the bf16 base at (measured-)identical healed quality.
Limitations
Thinking mode strongly recommended. The no-think fast path shows
larger quality loss (inherited from ternarization).
A knowledge gap vs the bf16 base remains on MMLU/ARC (~10–20 points)
— this is ternarization damage, not a reasoning defect.
German instruction-following was trained/evaluated on a proprietary
internal suite; the public benchmarks above are English.
BitNet-b1.58-2B could not be measured in the identical harness (its
i2_s GGUF requires the BitNet fork); we do not quote third-party
numbers in this table.
Speculative decoding with this model as a draft for the bf16 base is
blocked by the trimmed vocabulary (151669 vs 151936) in stock
llama.cpp — vocabulary-padded export is future work.
The published GGUFs quantize the (frozen, never-trained) embedding
and output tensors on the identical ternary grid as the upstream
pack; behavioral identity to the full-precision-embedding variant
was verified on held-out (identical scores, same environment).
Lineage & method
Qwen3-1.7B (Apache 2.0) → ternary packing by prism-ml (Apache 2.0) →
QAT healing + doubt-cut distillation (this work). Training data:
~140 verified teacher traces + ~200 self-repair pairs mined from
official train splits (GSM8K/ARC/OpenBookQA train — zero eval-set
contact, text-overlap verified). Full research log available on request.
