@misc{wong2026-malay-model-family-notes,
  author = {Wong, Sam},
  title = {Oaica 35B-A3B Malay v1.0 260923 model family notes},
  year = {2026},
  howpublished = {Oaica Research},
  url = {https://research.oaica.com/2026/10/malay-model-family-notes/}
}

Oaica 35B-A3B Malay v1.0 260923 model family notes

Sam Wong · sam.wong@oaica.com · Oaica (oaica.com) · 2 October 2026

These notes describe the general and safety variants and two experimental specialists. The family is fine-tuned from the open-weight base Qwen3.6-35B-A3B. The figures are local measurements, with protocol and provenance limits. None establishes comparative deployment safety or readiness for unreviewed decisions. The safety variant and its base model were also run on an Indonesian SEA-HELM set, a language neither was tuned for, to check that the measurement issues hold across languages; those cross-language figures are reported in the paper and rank no model.

1. Candidate uses

The same weights are not tied to one accelerator class. Our .oqm serving engine can pin the model’s mixture-of-experts layers in host CPU RAM, filling the available GPU memory automatically and offloading the remainder, so the family runs on laptop- and desktop-class GPUs at reduced context as well as on datacenter servers. The weights ship in our own .oqm quantised container, and higher-throughput engine configurations are offered for API and cloud inference. Runtime and context memory still require workload-specific budgeting. Fully local operation requires external services and outbound integrations to be disabled. Checkpoint names in these notes follow the local records; immutable checkpoint-to-product mapping remains pending.

2. Knowledge and coding measurements

MalayMMLU below uses 484 scored questions, greedy generation with reasoning off and answer-letter extraction. It is a separate protocol from the paper’s historical 500-question, reasoning-on results of 82.8 and 83.4. HumanEval uses 164 coding problems. Figures describe each test separately; percentages across different tests are not interchangeable. Counts are included to make the denominators explicit. No paired significance claim is made.

These are descriptive point estimates, not population guarantees. The stock reference’s 3–4 extracted correct letters out of 484 are a format/extraction failure; we exclude them from capability comparisons. The paper reports 80.7 for a separate full-set first-token run of that reference, which is also not comparable to the sample figures here. No 80-point Malay capability uplift or recipe superiority is claimed.

3. Recomputed safety measurements

Each run uses 1,572 scored items and our reproduction of the SEA-HELM Malay balanced-accuracy formula: toxicity plus three cultural subtasks. Greedy generation, reasoning off, batch size 8. The intervals below describe conditional item-sampling uncertainty, retaining shared prompt/response structure; they exclude generation reruns, training and model selection.

The unreconciled specialist-panel values are not included here. The historical safety-variant 47.2 and general-variant 45.6 belong to different runs and are not inserted into this panel. No safety ranking follows from overlapping marginal intervals; a paired comparison would be needed. The evidence document provides the counts, run hashes and score-recomputation code and identifies missing provenance.

4. Specialist limits

On HumanEval Oaica 35B-A3B Malay Coder v1.0 260923 passed 77 of 164 problems, compared with 126 for the stock reference. A separate LiveCodeBench v6 run reported 28.6% on 175 problems; the different benchmark does not itself contradict HumanEval. On a small, easy 16-item Malay-instruction coding set, it passed all 16; this is a feasibility observation, not broad validation. Oaica 35B-A3B Malay Researcher v1.0 260923 passed 129 of 164 HumanEval problems, but that alone does not validate research ability. Neither specialist has completed the domain, safety and over-refusal evaluation needed for a deployment recommendation. Long-context evaluation remains incomplete.

5. Permitted claims

Use measured values with their protocol, denominator, run and limitations. Describe local deployment as a configuration to validate. Describe moderation and drafting as supervised pilot candidates. Do not claim safety superiority, regulatory approval, verified absence of contamination, a universal capability ceiling or replacement of human reviewers.

Research: research@oaica.com. General enquiries: info@oaica.com. Full paper. benchlint checks reporting hygiene; it does not certify a model.