Strands Decider 2B¶

Adapter: strands-decider — Qwen3.5-2B torso + pointer head
License: Apache-2.0
Parameters: 2B · Architecture: Qwen3.5-2B-Base + LoRA + pointer head
Languages: multilingual · Context: 4,096 tokens

Default variant¶


Repo	`onnx-community/strands-decider-2B-hobson-v19-ONNX`
Files	`onnx/model_q4f16.onnx` + `.onnx_data`, `tokenizer.json`, `config.json`
Quantization	4-bit weights, fp16 activations
Download	1.11 GB
Role	default

Variants¶

Variant	Quantization	Size	Role	Install
`q4f16` *	4-bit + fp16	1.11 GB	default	`dtk models download strands-decider-2b --variant q4f16`
`q8`	int8 block	1.82 GB	quality	`dtk models download strands-decider-2b --variant q8`

q8 is the closer-to-reference build; q4f16 is smaller and faster on CPU. Select one with dtk models use strands-decider-2b --variant <name>. See Variants and quantization.

Reproducibility¶

Revision: 31a83e0244b26d939ae33f87189d6c8204003e53
Default file hashes (SHA256):
onnx/model_q4f16.onnx — 8ee20666f274ebb3774bc98e0026f3bab9cffd3796440c73d7700169e756b68a
onnx/model_q4f16.onnx_data — d442842f1c5db3dd5b84ac35f683b408a1330075dae6f8f51dcdc5154b64f920
tokenizer.json — a2cdd2e108566b09079afa8d266e9e65e7c280f27b771218237a06de5ba9cd86
config.json — 707931b0fef6a8581964f6dc5f6ae97ba8cab078cfb5c9cfff958272e8146343
Verify: dtk models verify strands-decider-2b --variant q4f16
Bench: uv run python bench/run.py --suite bench/suites/smoke.jsonl --models strands-decider-2b
The q8 hashes live in catalog.py.

Evidence (2026-10-21)¶

Reference fixture from the upstream export (Help! My payouts have been failing for 3 days!):

Variant	urgent P(true)	team billing	frustration score
upstream reference	0.829	0.844	1.10
`q4f16`	0.808	0.845	1.11
`q8`	0.816	0.836	1.10

Max probability delta on the three reference questions: 0.018 for q4f16, smaller for q8 on most fields.
CPU p50 after warmup: 719 ms q4f16, 1,467 ms q8.
The upstream card reports q4f16 max |Δp| 0.072 on 11 questions and q8 0.016 on its WebGPU checks.

Contract notes¶

Prompt: <state>…</state> plus one <question type="…"> block ending in <answer>; options are numbered and each option's last token is a pointer.
Graph inputs: input_ids, attention_mask, answer_pos, option_pos.
Logits are divided by decider.temperature_by_kind and softmaxed per question.
noul uses false/true slots; score uses level indices with descriptions.

Known limitations¶

2B model: CPU latency is seconds, not milliseconds; q8 is slower than q4f16.
The upstream project warns that behavior under concurrent requests is not verified and that its server is experimental.
The ONNX export is a WebGPU-oriented quantized build; use q8 when probabilities need to be closest to the reference.
No image/audio support in this text-only projection.

Loading each variant¶

REST selects the quantization through the standard model field using the alias strands-decider-2b@variant; there is no variant field in the request body.

`q4f16` (q4/fp16, default)¶

Library

from decision_toolkit import DecisionModel

model = DecisionModel("strands-decider-2b", variant="q4f16")

CLI

dtk models download strands-decider-2b --variant q4f16
dtk models use strands-decider-2b --variant q4f16
dtk decide --model strands-decider-2b --variant q4f16 \
  --state "..." --q 'urgent=noul:Is this urgent?'

REST

curl -s http://127.0.0.1:8000/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{"model": "strands-decider-2b@q4f16", "state": "...",
       "questions": {"urgent": {"type": "noul",
                     "instructions": "Is this urgent?"}}}}'

`q8` (int8, quality)¶