Runtime: Osaurus 0.25.26 or newer.

OsaurusAI/d3-turbo-JANGH2

d3-turbo in 14.5 GB: the same decision as the full-precision model on 92.65 % of 2,000 requests of tasksource/tasksource-jev-typed-decisions test.

A JANG bundle of vllm-sr/d3-turbo, the upgraded 27B multimodal decision model of vLLM Semantic Router's Decision 3.0. Give it a state (text or JSON, optionally with images and videos) and typed questions: pick one of several options, yes or no, or a score on an ordered scale, or where something is (one box with its confidence, in an image or at a moment of a video). It answers every question with a probability per option in one forward pass and generates no text.

• Layout: MLP JANGH 2-bit (Hadamard-32 rotation, odd-cubic codebook) with the most sensitive layers higher; attention, linear attention, embeddings and vision tower affine 8-bit. The 255-way answer head (decision/readout.safetensors) is the original FP32 tensor, byte for byte; the linear-attention gate inputs stay BF16.
• Size: 14.50 GB of model shards, 4.45 bits per weight over the 26.09B source parameters (BF16: 52.2 GB).

Agreement with the full-precision model

Each question is answered by this bundle and by the BF16 model; decisions counts the same top option (or the same side of 0.5 for yes / no), KL is measured on each question's option probabilities. None of these questions were used for calibration.

test set	questions	decisions	mean KL ↓	gold accuracy, BF16 → this bundle
Decision Index 0.3 public suite, 466 of its questions across the 41 benchmarks	466	91.63 %	0.0592	76.4 → 74.5 %
LocalLLaMA/typed-decisions test (800 questions, soft gold)	800	93.12 %	0.0262	73.0 → 76.1 %
held-out decisions over our own texts, images and videos (487 questions)	487	93.22 %	0.0285	—
2,000 requests of tasksource/tasksource-jev-typed-decisions test (the split the vendor scored its quantized packages on)	2,000	92.65 %	0.0466	83.5 → 82.5 %

Box questions: on 29 held-out Box questions over code, charts, documents, screenshots, terminals and an invoice (small targets that the full-precision model itself answers with low confidence) boxes shift under any perturbation, so agreement there is measured against that floor: this bundle's box overlaps the BF16 box at a mean IoU of 0.40 (JANG_6M, near-lossless on decisions: 0.52).

All three bundles

bundle	size	decisions (tsj, 2,000)	decisions (suite)	mean KL (suite)
d3-turbo-JANG_6M	23.6 GB	99.25 %	98.07 %	0.0006
d3-turbo-JANG_4M	19.3 GB	97.65 %	97.42 %	0.0052
d3-turbo-JANGH2	14.5 GB	92.65 %	91.63 %	0.0592

How it was made

Calibration ran d3-turbo's own decision forward (prefix-bidirectional attention with the state read once, the decision prompt format, images and videos) over 2,899 decision prompts and 1.96M tokens: routing questions over real requests, tool selection over real tool lists with up to 255 options, agent and terminal traces, code, security, long documents, typed decisions, images and screen videos, and 121 Box questions over the calibration images. Activation-aware scales (AWQ) are folded into the producing norms and projections, and every quantized MLP is rounded with GPTQ on Hessians from that forward, each module kept only where it beats plain rounding.

License

Apache-2.0, as the base model vllm-sr/d3-turbo (built on Qwen/Qwen3.8-27B). NOTICE and LICENSE are carried over. Quantized by Osaurus (eric@osaurus.ai).