CLM v0.1 8B, compiled for the SiMa.ai MLSoC Modalix
CLM v0.1 8B by Contrastive-LM is a "System-1" decision model, a Contrastive Language Model: a frozen Qwen3-8B embeds the state and each option, and two small projection heads compare them, to pick one of several options, score on a scale, or say yes or no.
This is that checkpoint compiled for the Modalix MLA. The encoder is a chain of 18 graphs of
two layers each, with int8 weights and bfloat16 activations (A_BF16_W_INT8, 7.25 GB on
the MLA), for rows of up to 128 tokens. One pass through them (254 ms for 128 tokens on a Modalix
DevKit, inside the runtime) holds a question's state and all its options: about 0.3 s for a
new question, nothing for one seen before. With int8 weights it is an approximation of the
PyTorch model (embedding cosine about 0.99; 9 of 13 reference decisions the same), and a
one-token option is embedded poorly. It is 9.12 GB to download.
What is in the folder
| file | what it is |
|---|---|
qwen_s128_l<NN>n2_stage1_mla.elf |
two layers of the encoder, from layer NN: 18 graphs that run one after another |
token_embeddings.bf16 |
Qwen3-8B's embedding table, looked up on the CPU |
heads.f32, heads.json |
CLM's state and action heads, run on the CPU, and their layout |
tokenizer.json |
Qwen3-8B's tokenizer |
clm_config.json |
which graphs there are, in order |
Using it
The files run on a Modalix board with the runtime and web app from neat-decision-studio, whose Models page downloads the model from here onto the board and loads it on the MLA; the list it reads is in dotimothy/sima-laya, with the Laya models. By hand:
hf download dotimothy/sima-clm --include "clm/*" --local-dir .
./laya run clm --state "We were billed twice for March." \
--question '{"type": "noul", "instructions": "Is this a billing problem?"}'
It is not usable with PyTorch or ONNX Runtime; for that, use the original checkpoints, Contrastive-LM/CLM-v0.1-8B and Qwen/Qwen3-8B.
License and attribution
Apache-2.0. clm holds the projection heads of Contrastive-LM's CLM-v0.1-8B
(https://github.com/Contrastive-LM/CLM, Apache-2.0), unchanged, and Alibaba Cloud's Qwen3-8B
(https://huggingface.co/Qwen/Qwen3-8B, Apache-2.0): its embedding table and tokenizer
unchanged, and its 36 decoder layers compiled for the MLA with their weights rounded to int8
and no language-model head.