BranchOne returns Choice probabilities from Qwen3.5-0.8B-Base, median 0.305 s on ten short text questions

Necati Demir posted BranchOne on October 29, 2026. The open checkpoint returns one probability per supplied candidate from Qwen3.5-0.8B-Base plus LoRA. His A100 Colab run prints median 0.305 s on ten short text questions. Apache-2.0. This desk did not run the weights.

Necati Demir posted on October 29, 2026, at 13:31 UTC, that Jev made System One decision models visible and that he built an open model in that category named BranchOne. The link is his BranchOne article. On October 30 he wrote that more choices mean more computation and that the inference path repeats the common text prefix.

What it returns

The article and the Hugging Face card ndemirdemir/BranchOne, created October 23, 2026 and opened October 31, describe a Choice path. The caller supplies state, a question, and 2 to 150 candidate descriptions, optionally one image. The model returns selected_id and one probability per candidate. The backbone is Qwen/Qwen3.5-0.8B-Base with rank-8 LoRA and a shared scoring head. License Apache-2.0. GitHub ndemir/BranchOne holds training and inference code.

What the author timed

On an NVIDIA A100-SXM4-80GB Colab run printed in the article, model loading took 7.884 s, warm-up 2.300 s, and ten short text predictions 3.056 s. Mean per prediction was 0.306 s, median 0.305 s, range 0.304 to 0.308 s. Nine of the ten illustrative selections matched his intended answers in that run. The stock-rule example selected the wrong stop-selling option at 35.6%. The card says those ten questions are not a representative accuracy benchmark. No hosted POST on the pages this desk opened. This desk did not run the weights.

0.305 s is that Colab median. It is not a JevBench composite and not a Decision Index Full score. OR-decision-4B and MiniCPM5-2B-Jev keep their own JevBench cells on /news/or-decision-4b/ and /news/minicpm5-2b-jev/.

This site's reading

Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.

Verify

Necati Demir, @ndemir, October 29, 2026, 13:31 UTC, status 2115798561640030268. The text says Jev made System One decision models visible, that those models choose among supplied options and return probabilities, and that he built an open model in that category named BranchOne. The link is n.demir.io/articles/branchone-system-one-decision-model/. On October 30, 13:31 UTC, status 2116161068628074586, he wrote that more choices mean more computation and that the inference path repeats the common text prefix.

The article and Hugging Face card ndemirdemir/BranchOne, created October 23, 2026 and opened October 31. Apache-2.0. Backbone Qwen/Qwen3.5-0.8B-Base with rank-8 LoRA and a shared scoring head. Input is state, question, and 2 to 150 candidates, optionally one image. Output is selected_id and per-candidate probabilities. The author's A100-SXM4-80GB Colab timing table prints model loading 7.884 s, warm-up 2.300 s, ten predictions 3.056 s, mean 0.306 s, median 0.305 s, range 0.304 to 0.308 s. Nine of ten illustrative selections matched his intended answers in that run; the stock-rule example selected the wrong stop-selling option at 35.6%. The card says those ten questions are not a representative accuracy benchmark. GitHub ndemir/BranchOne holds training and inference code. No hosted POST on the pages this desk opened. This desk did not run the weights.

Compare

0.305 s is the author's median on ten short text questions with three choices each on an A100. It is not a board latency rank and not a desk rerun. Celeris-1 Decision's printed one-question median is 67 ms. hunch's published CPU clocks are 20 ms on text and 100 ms on images. Do not merge those three.

BranchOne has no published JevBench composite and no Decision Index Full score on the pages this desk opened. OR-decision-4B's card JevBench 83.55% and MiniCPM5-2B-Jev's 78.79% stay on their own pages. Jeva-4B's 0 flips and BANKING77 91.62 stay on /news/jeva/.

The nine-of-ten match in the Colab write-up is the author's illustrative set. It is not a board accuracy.

Terms

BranchOne: Necati Demir's October 2026 open System One checkpoint. Qwen3.5-0.8B-Base plus LoRA and a shared scoring head return one probability per supplied candidate. Apache-2.0. Hugging Face ndemirdemir/BranchOne. His A100 Colab median on ten short text questions is 0.305 s.