Wringer p3b_w2 — Agents-A1-4B at 2.655 bits/weight

Draft v0.2 (2026-09-12). HumanEval is reported as the mean ± sd over 4 sampling seeds (temperature 1.0); IFEval and GSM8K are single official runs.

A 2.655 bits-per-weight (body) quantization of InternScience/Agents-A1-4B (Qwen3.5 hybrid: 25 GatedDeltaNet linear-attention layers + 7 full-attention layers, 3.57 B quantized body weights), produced with Wringer: fixed-grid GPTQ codes → one round of low-rank "water" (r128 LoRA, KD to the bf16 parent on the model's own long-trajectory corpus) → closed-form "wring" that re-solves scales and codes so no adapter is shipped. Two fill-and-wring rounds were run.

Scores (official A1 evaluation framework, thinking on, 16k max tokens)

Model	body b/w	IFEval (prompt, strict)	HumanEval (4-seed mean ± sd)	GSM8K	comp*
bf16 parent (anchor)	16	92.79	93.75 ± 1.52	95.53	1.000
this model (materialized from the container = what you download)	2.655	88.17	85.37 ± 1.32	93.48	0.9465
this model, research-state weights (fp32 scales)	2.655	86.32	86.74 ± 2.74	92.49	0.9412
single-round control p3b_w	2.655	87.80	85.37 ± 2.63	92.95	0.9433
zero-training p3b (same codes, no water)	2.655	75.97	70.73 (1 seed)	83.70	0.8144
llama.cpp GGUF at the same body b/w (U27, IQ2_M-class, 2.763)	2.763	73.38	63.41	76.65	0.7547

*comp = mean of the three retention ratios against the bf16 anchor (HumanEval ratio uses the 4-seed means). The three Wringer rows are within sampling noise of each other (HumanEval single runs swing by up to 5.5 pp at temperature 1.0); the second wring round did not hurt and is what we publish. Body b/w counts the quantized linear layers only; embeddings, norms and lm_head stay bf16 (whole-LM figure 4.69 b/w with bf16 embeddings; 2.95 b/w if the tied embedding were stored at Q4_K-class 4.5 b/w — not evaluated).

HumanEval with the 10 tasks flagged by our contamination check removed (see below), seed 20260806: 89.61 (research state), 94.16 (anchor).

What is in this repo

wringer_p3b_w2.safetensors — the container (1.10 GiB): packed codes + fp16 / int8 block scales, metadata wringer_meta.
wringer_unpack.py — dependency-free materializer: container + bf16 source export → bf16 HF checkpoint.
bf16/ — the materialized bf16 checkpoint (fake-quant weights; loads with transformers / vLLM like the parent).
reports/ — packing verification (pack_p3b_w2.md), contamination check (contamination_e69.md), verdict JSON.

No custom kernel is provided. The container is a storage format; inference uses the materialized bf16 weights.

Recipe (per module)

Module	grid	block	scales
self_attn.k_proj, v_proj	int8	per row	fp16
self_attn.o_proj	16-level	128	int8 + fp16 row scale
linear_attn.in_proj_z, in_proj_qkv, mlp.gate_proj, linear_attn.out_proj	4-level {−2..1}	128	int8 + fp16 row scale
self_attn.q_proj, mlp.up_proj, mlp.down_proj	8-level {−4..3}	128	int8 + fp16 row scale

Ledger: code 2.588 + scales 0.067 = 2.655 b/w. Codes solved by GPTQ per layer on 128 × 16k self-generated calibration rows; scales by a joint least-squares closed form with a prior (λ = 0.01).

Honest caveats

Scale precision. The research solver stored scales in fp32 while the ledger charged fp16. Materializing at ledger precision moves 2.95 % of the bf16 weights by one ulp. We re-scored the container-materialized weights: HumanEval 85.37 ± 1.32 vs 86.74 ± 2.74 for the research state (difference 1.37 pp, SE 1.52 — not distinguishable), IFEval and GSM8K slightly higher. What you download is exactly what the first row measures. The solver now projects scales to fp16 at solve time.
Sampling noise. HumanEval at temperature 1.0 swings by several points between seeds, even for the bf16 parent (91.46–94.51). Single-run HumanEval headlines from this project's earlier write-ups (e.g. 89.02) were lucky draws; all HumanEval numbers here are 4-seed means.
Contamination. Our calibration corpus is model-generated. An 8-gram check against the three test sets found GSM8K and IFEval clean; 10 HumanEval tasks have ≥25-token overlaps with corpus solutions (list in reports/). Scores are reported with and without them.
Three benchmarks only. IFEval / HumanEval / GSM8K in thinking mode. No MMLU-class or agentic suites yet.
One model, one architecture. Generality beyond this hybrid Qwen3.5 model is not yet shown.