laya-browser-v32b (v32b-b15)

A 322M-parameter browser-agent decision model, fine-tuned from cklxx/laya-browser (which fine-tuned convaiinnovations/laya for browser decisions). Fully Apache-2.0 upstream; this checkpoint keeps the same license.

What it is: a System-1 decision head for browser agents. It never writes text — it answers typed questions about a page state (which operation? which element? does the goal-holding statement hold?) and returns calibrated probabilities. That makes it safe to put in front of a browser: it can only choose from what it was shown.

Headline numbers (all measured 2026-09-25, same fixtures for every model):

Benchmark	v32b-b15	official laya-browser (td)	hosted Jev API
recovery2-holdout (240)	0.7125	0.425	—
browser-suite v5 (110)	0.5636	0.5818	—
browser-suite v4 (70)	0.5143	0.500	—
MiniWoB-116	0.9138	0.6638	—
JevBench hard (111)	0.4144	0.243	0.7207
JevBench easy (48)	0.8542	0.979	1.0000
decision latency (p50)	27 ms (RTX 3080)	—	854 ms (network)

vs the official browser-tuned checkpoint: wins holdout (+28.8pp), MiniWoB (+25pp), suite v4 (+1.4pp), JevBench hard (+17.1pp), injection_safety (0.75 vs 0.50); loses suite v5 by 2 items. vs the hosted Jev API: lower absolute accuracy but $0 cost, fully private, offline, 31x faster — and it beats Jev on score questions (0.667 vs 0.333) and temporal_numeric (0.33 vs 0.20) on JevBench hard.

What was trained

Starting from the official v10s-lineage browser checkpoint, this model adds (v17→v32 recipe, frozen encoder + head-only training throughout):

1. SCROLL_UP + recovery data — the official checkpoint had a 0/203 hit rate on scroll-up; this one is 193/203 with no loss of general accuracy.
2. Counterfactual ranking pairs — Sharper target choice among similar candidates.
3. Noul (statement-holds) training — the upstream training pipeline never produced noul items (yes/no judgment questions), leaving done_judgment at 0.615. A purpose-built 8k-item noul corpus (8 patterns × 320, balanced yes/no) raised probe accuracy from 37% → 100% and done_judgment to 0.769 (probability-averaged with the choice specialist).
4. Head blending — capability heads are blended into the champion at low weight (6–15%) so new skills arrive with zero regression elsewhere.

The encoder (134 tensors) is byte-identical to upstream; only the 36-tensor decision head was trained.

Usage

Requires laya (pip install laya) or laya-mlx on Apple Silicon.

import laya

agent = laya.load("ichenney/laya-browser-v32b", subfolder="v32b", device="cuda")
state = {...}  # page state: url, title, text, elements table
questions = {
    "operation": {"type": "choice",
                  "instructions": "Pick the operation that achieves the goal.",
                  "criteria": {"0": "CLICK", "1": "TYPE_TEXT", "2": "SCROLL_DOWN",
                               "3": "SCROLL_UP", "4": "SELECT_OPTION", "5": "WAIT",
                               "6": "DONE", "7": "BLOCKED"}},
}
answers = agent.predict(state, questions)["answers"]
print(answers["operation"]["choice"], answers["operation"]["probabilities"])

Or through the localdecide harness (recommended — it adds element-table building, coarse-to-fine chunking, and guards). v32b is the harness's default model — model="browser" loads this repo automatically:

from localdecide import BrowserDecider
from localdecide.backends import LayaTorchBackend

backend = LayaTorchBackend()          # model="browser" → this checkpoint
# to stay on the upstream cklxx v10s instead:
# backend = LayaTorchBackend(model="browser-legacy")
decider = BrowserDecider(backend=backend)

Checkpoint contents

v32b/
├── model.safetensors      # full model: frozen encoder (134 tensors) + trained head (36)
├── rl_agent_config.json   # laya agent config (max_len 1024, head_max_len 768, v1 format)
├── encoder/config.json    # ModernBERT-style encoder architecture config
└── tokenizer/             # tokenizer.json + config (same vocab as upstream)

Training details

• Base: cklxx/laya-browser (frozen encoder — 134 tensors untouched, byte-identical), head-only fine-tune (36 tensors)
• Data (each family generated synthetically from real page fixtures, no user data):

  corpus	items	purpose
  general browser decisions (v10s-lineage mix)	~11,800	retain upstream competence
  targeted 6-family patch (op-choice / state-reason / multistep / surface-trap / injection / filter-first, 8 languages)	9,600	v20: fix known failure axes
  counterfactual ranking pairs	11,200	v21: sharper target choice (blend-only)
  SCROLL_UP + recovery	6,972	v23: upstream had 0/203 scroll-up probe
  multistep replanning	5,600	v25: proved harmful, archived
  22-failure-mode patch	6,300	v29: evidence for the noul investigation
  noul statement-holds	8,000 (3,200 yes / 4,800 no)	v31: the root-cause fix — 8 patterns × 320, no-heavy to counter the model's yes-bias (probe NOUL_NO was 37%)

• Optimization: head LR 1e-4, bf16 autocast, batch 8, GRPO-style grouped advantages (group 4, w_sph 0.75 / w_rps 1.0), 1–3 epochs per stage
• Cost: ~3 minutes per training run on one RTX 3080; the entire 15-version campaign totals under 2 GPU-hours
• Method: single-variable iterations — change one thing, run the full eval battery (browser-suite v4/v5 + recovery2-holdout, ~90 min), keep or discard on data. New capabilities are trained as separate heads and blended into the champion at 6–15% weight, which transfers skills with zero measured regression.
• Champion selection: highest held-out mean across the 3 suites; the final ensemble (v36) probability-averages v32b-b15 with the v31a noul specialist at w=0.4
• Evaluation: 3 suites × 3 official baselines + hosted Jev API (fresh 231-item run), all JSONs published in the training log

Limitations

• 322M parameters: complex multi-hop reasoning (JevBench multi_hop 0.22 vs hosted Jev 0.89) is beyond current capacity — this is a capacity, not a recipe, gap
• English-dominant training data (the synthetic fixtures cover 8 languages, but natural-page diversity is English-heavy)
• Safety-adjacent judgments (phishing, abuse) should be guarded by harness rules, not trusted to the head alone — the v10s phishing analysis in the harness README applies to v32b too
• FP32 weights: 1.3 GB on disk. An FP16/BF16 export would halve that; the MLX runtime loads it fine on 8 GB Apple Silicon but the torch path wants ~2 GB resident

License

Apache-2.0 (inherited from both upstream checkpoints). The companion harness is github.com/ChenneyZhuang/laya-browser-agent.

Citation

@misc{laya-browser-v32b,
  title  = {laya-browser-v32b: a noul-capable browser decision model},
  author = {Chenney Zhuang},
  year   = {2026},
  url    = {https://huggingface.co/ichenney/laya-browser-v32b},
  note   = {Fine-tuned from cklxx/laya-browser (Apache-2.0), frozen-encoder head training with noul corpus injection}
}