Jev directory
CommunityJev Models & Reimplementations
Strands-Decider 2B Hobson v19 ONNX
A browser-ready ONNX conversion of Strands Labs' strands-decider-2B-hobson-v19 by onnx-community: q8 (1.80 GB) and q4f16 (1.09 GB) builds that return calibrated option probabilities, scoring a card-reported 59.4% on typed-decisions via WebGPU.
- Reading time
- 2 min
- Published by
- Community
- Author
- ONNX Community
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Platforms
- Browser
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- ONNX conversion of StrandsAgents/strands-decider-2B-hobson-v19: state plus typed choice, score and noul questions in, calibrated probability per option out.
- Two dtypes: q8 (1.80 GB, 8-bit block-32 decoder with 4-bit embeddings) and q4f16 (1.09 GB), both needing WebGPU with shader-f16.
- Runs in the browser through open-jev with model: "strands-decider-2b", or with ONNX Runtime anywhere.
- Versus the PyTorch fp32 upstream: q8 max probability difference 0.016–0.026 with 0 top-answer changes; q4f16 0.072–0.146 and 5 changes on long states.
- Not trained on typed-decisions yet scores 59.4% overall there with q8 (337/600 choice, 458/800 score, 392/600 noul); answers at 0.9 or more are right 94%.
Watch out
Card reviewed 2026-10-17; the typed-decisions accuracy (58.4% q4f16, 59.4% q8), the 94% confidence figure and the M3 Pro timing are self-reported on the card, and the model was not trained on typed-decisions. q4f16 moved near-ties on long states (5 top-answer changes), so the card prefers q8 unless download size matters more.
Questions about this entry
Short answers drawn from the linked project or page.
A browser-ready ONNX conversion by onnx-community of StrandsAgents/strands-decider-2B-hobson-v19, Strands Labs' decision model. Given a state and typed questions (choice, score, noul), it returns a calibrated probability for every option in one forward pass per question.
It keeps the decision-model shape: state plus typed questions in, probabilities over declared options out, with no generated answer text. The card points to the original model card and the strands-decider repository for training data, evaluation and limits.
Two Transformers.js dtypes: q8 as onnx/model_quantized.onnx (1.80 GB, decoder 8-bit with block 32 and 4-bit embeddings) and q4f16 as onnx/model_q4f16.onnx (1.09 GB, decoder and embeddings 4-bit).
Through open-jev with model 'strands-decider-2b': on a MacBook Pro M3 Pro in Chrome with WebGPU it took about 130 ms for one short question, with questions on one state batched. ONNX Runtime on CPU runs the files too; both need WebGPU shader-f16 in the browser.
Against the upstream PyTorch fp32 engine, the largest option-probability difference was 0.0009–0.0013 for fp32 ONNX, 0.016–0.026 for q8 on WebGPU and 0.072–0.146 for q4f16, with 0 top-answer changes except 5 for q4f16 on long states.
Not trained on that set, q8 reached 59.4% overall (337/600 choice, 458/800 score, 392/600 noul) and q4f16 58.4%, with a median of 2.2 s per state; answers at 0.9 confidence or more were right 94% of the time.
The rank-16 LoRA (alpha/r = 2) is merged into Qwen3.5-2B-Base, the LM head is dropped, and embeddings, decoder and a roughly one-million-parameter pointer head are fused into one graph; swapped weights keep cosine similarity of at least 0.993. Licence: Apache-2.0.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Text-to-screenshot (Malay Vasa)
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
X: Google's MediaPipe Decision Maker, tested in the browser
More Jev resources like this
prithivMLmods' GGUF builds of autotrust/JEV-35B, the Qwen3.5-35B-A3B MoE decision model that returns calibrated choice, noul and score answers over text and images; files run from BF16 to Q3_K_L for llama.cpp.
Jev Models & Reimplementations#community#open-models#local
Community
A calibrated image-capable 27B decision model that fits an 8 GB GPU: the JEV LoRA-and-head recipe trained on Prism ML's ternary Bonsai 2, returning typed probabilities with no generated text.
Jev Models & Reimplementations#community#open-models#local
Community
A community Q8_0 GGUF of Mapika/decider-0.8b, the smallest System One typed-decision model: one forward pass returns calibrated Choice, Score and noul answers, reporting 0.405 on JevBench hard-public at 920 MB resident.
Jev Models & Reimplementations#community#open-models#local
Community
A 4-bit AWQ plus GPTQ build of AutoTrust's JEV-35B from AI Outfitters: INT4 weights with FP8 and INT8 activations, calibrated on all 256 experts, and served by vLLM on a single 24 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
A 4-bit W4A8 AWQ plus GPTQ build of AutoTrust's JEV-27B-VL from AI Outfitters that keeps the rank-32 LoRA decision head and a bf16 vision tower, served by vLLM on a single 24 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
mradermacher's static GGUF quantizations of IqSantiagow's 0.8B edceleste decider for llama.cpp, from Q2_K (0.5 GB) to f16 (1.6 GB), with Q4_K_M and Q4_K_S marked fast and recommended; no imatrix quants are planned by this quanter.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.