Contrastive Language Models score a state against a set of candidate actions with a contrastive objective. Two projection heads on frozen Qwen3-8B embeddings, trained with InfoNCE; clm-serve maps Choice, Score and Noul onto candidate ranking.
Standard Thinking's decision model. Ministral 3 8B with a merged LoRA, served through stock SGLang and an open jev-adapter that exposes /v1/systemone and scores the supplied options in one forward pass with per-type temperatures.
Decides
choice, score, noul, rank, classify, route
choice, score, noul, classify, route
Architecture
clm
standard-one
Fine-tuned from
qwen/qwen3-8b
mistralai/ministral-3-8b-instruct-2512-bf16
License
apache-2.0
apache-2.0
Availability
Open weights
Open weights + hosted API
Hosted by
—
Standard Thinking
Input price
—
—
Decision accuracy
—
71.1%
Calibration error
—
—
Valid action rate
—
—
Median latency
—
25.8 ms
p95 latency
—
41.9 ms
Figures are from each model’s manifest; accuracy and latency are what the publishers report, on their own suites and hardware. Add a third model.
Questions
What is the difference between clm and standard-one?
clm is from Contrastive-LM and standard-one from Standard Thinking. clm has open weights you can download and run; standard-one has open weights and a hosted API. Both answer choice, score, noul, classify and route questions. Only clm answers rank.
Which is more accurate, clm or standard-one?
Only standard-one publishes an accuracy figure (71.1% on typed-decisions suite, 400 cases / 2,000 decisions (maker's served run, v2)); clm does not, so there is no comparison to make without your own test.
Which is cheaper, clm or standard-one?
clm: Free (open weights). standard-one: Hosted, price not published, or free to self-host. Open weights cost nothing per call beyond your own hardware.
Can I run clm or standard-one locally?
Yes, both: systemone pull contrastive-lm/clm and systemone pull standard-thinking/standard-one download the weights.
