70.14% four-panel mean
across 2,720 decisions. Results include strengths and remaining gaps across all 27 tasks.
Model
Mean accuracy ↑
Decisions
Composition
Reading
Inference
Sol · 2B · v1.3
70.14
73.75
46.08
76.56
84.17
Nox · 4B · v1.3
75.03
83.00
51.79
79.06
86.25
Kev · 9B
73.09
76.36
45.54
86.72
83.75
Decider · 2B
71.75
64.01
46.58
92.03
84.38
Qwen3.5 · 4B · untuned
70.25
69.89
43.33
87.97
79.79
Qwen3.5 · 2B · untuned
60.54
57.12
39.00
73.75
72.29
Laya · Upstream default
52.44
57.01
37.75
51.25
63.75
Jev · 1.13.0 · frontier
82.45
79.10
66.38
94.53
89.79
Accuracy (%). Each panel contributes one quarter with its frozen family/source weights.
Bold
marks Sol or Nox strictly above every external open reference for that metric; the other Decision model and Jev are excluded from the threshold. Jev is a closed-service frontier reference.
Methods and uncertainty
.
Detailed capabilities
Semantic consistency
Model
Consistent / eligible groups
Consistency ↑
Sol · 2B · v1.3
167/192
86.98
Nox · 4B · v1.3
174/192
90.62
Kev · 9B
143/192
74.48
Decider · 2B
126/192
65.62
Qwen3.5 · 4B · untuned
98/192
51.04
Qwen3.5 · 2B · untuned
94/192
48.96
Laya · Upstream default
115/192
59.90
Jev · 1.13.0 · frontier
158/192
82.29
Mixed equivalent variants of language, keys and presentation;
consistency is not accuracy
. A consistently wrong answer still counts. This observed-core supplement weights 192 eligible semantic groups equally and is outside the four-panel mean.
Definitions and exact counts
.
More questions, measured
Same inputs and physical AMD gfx942 GPU; 30 measured requests per point across three blocks. Python request latency includes tokenization and inference, excluding loading and network.
p95 and all three native types
.
Try it
Download
hf download llm-semantic-router/Decision-1.0-Sol --revision v1.3 --local-dir decision-model
, then follow
ROCm setup
. In that container, with the model mounted at
/model
:
from decision import DecisionModel
from decision.example import REQUEST
model = DecisionModel.from_pretrained("/model", local_files_only=True)
print(model.decide(**REQUEST)["answers"])
The complete state, question and candidates must fit 16,384 tokens; overflow is rejected. The bundled normalization profile loads automatically. AMD gfx942 is validated; CPU/MPS are unsupported and NVIDIA is unqualified. Use a fresh Python process when switching profiles.
Architecture
A causal Qwen3.5 text backbone combines gated linear and full attention. A shared candidate head reads candidate endpoints and the final query vector. Each question uses one forward pass; questions run independently in batches of eight.
Decision-1.0-Sol huggingface.co is an AI model on huggingface.co that provides Decision-1.0-Sol's model effect (), which can be used instantly with this llm-semantic-router Decision-1.0-Sol model. huggingface.co supports a free trial of the Decision-1.0-Sol model, and also provides paid use of the Decision-1.0-Sol. Support call Decision-1.0-Sol model through api, including Node.js, Python, http.
Decision-1.0-Sol huggingface.co is an online trial and call api platform, which integrates Decision-1.0-Sol's modeling effects, including api services, and provides a free online trial of Decision-1.0-Sol, you can try Decision-1.0-Sol online for free by clicking the link below.
llm-semantic-router Decision-1.0-Sol online free url in huggingface.co:
Decision-1.0-Sol is an open source model from GitHub that offers a free installation service, and any user can find Decision-1.0-Sol on GitHub to install. At the same time, huggingface.co provides the effect of Decision-1.0-Sol install, users can directly use Decision-1.0-Sol installed effect in huggingface.co for debugging and trial. It also supports api for free installation.