Oaica 35B Malay Agents 260827
Formerly: sprappcom/agents-a1-malay35b-dare-pqm
DARE-TIES merge of InternScience/Agents-A1 (agentic / AI-research focused) into
malay35B (Malay-focused CPT+SFT of Qwen3.6-35B-A3B), exported to .pqm for the
prism-engine inference server. Text-only (the donor's vision tower is stripped in
this merge).
License & attribution
This repository distributes a composite work under a dual arrangement:
-
Proprietary layer — © 2026 BCZ Singapore Pte Ltd. All rights reserved.
The .pqm container format and packaging, the DARE-TIES merge recipe, the Malay
continued-pretrain + SFT recipe, and the prism-engine inference server (Rust/CUDA)
are proprietary and are not licensed under Apache-2.0.
-
Base model weights — Apache-2.0. The underlying weights are a modified, merged
derivative of two Apache-2.0 upstreams:
| Component | Source | Copyright | License | Modification |
|---|
| Backbone | Qwen/Qwen3.6-35B-A3B | © 2026 Alibaba Cloud | Apache-2.0 | Malay CPT+SFT base; DARE-TIES-merged; vision tower stripped; weights repacked to .pqm |
| Agentic donor | InternScience/Agents-A1 | © InternScience | Apache-2.0 | DARE-TIES-merged (weight 0.5, density 0.5); weights repacked to .pqm |
These base weights remain licensed under
Apache-2.0. The full Apache-2.0 license
text and per-component attribution are in the
NOTICE
file in this repo.
This is not an original, trained-from-scratch foundation model — it is a modified, merged
derivative of the Apache-2.0 upstreams named above. Use of the base weights is governed by
Apache-2.0; use of the proprietary .pqm packaging, the merge/training recipe, and
prism-engine requires a separate license from BCZ Singapore Pte Ltd.
Requirements — read before downloading
- Inference engine:
.pqm is a proprietary container, not readable by vLLM,
llama.cpp, Ollama, or transformers. Runs only on prism-engine (proprietary
Rust/CUDA server, not included in this repo). Contact sprappcom for engine access.
- Tokenizer sidecar:
.pqm does not embed a tokenizer. Use the malay35b.tok
shipped in this repo via PRISMX_TOKENIZER (identical across all malay35B variants).
- Hardware: A100 (sm_80) / Ada-Ampere (sm_86/sm_89) validated. ~19 GB VRAM full
residency, or 8 GB with CPU-RAM expert offload.
Merge details
- Base: malay35B
- Donor: InternScience/Agents-A1 (vision tower stripped — text-only merge)
- Method: DARE-TIES, weight=0.5, density=0.5, normalized. The donor's per-expert
MoE tensors were stacked into the base's fused layout before merging; all 40 layers'
fused expert tensors (80 total) were merged, plus 611 attention/norm tensors.
Architecture
- Qwen3.6-35B-A3B MoE, hidden_size=2048, 40 layers (30 Gated DeltaNet + 10 full attention)
- 256 routed experts + 1 shared expert, top-8 routing
- vocab_size=248077
Quantization
| Tensor group | Type |
|---|
| token embedding, attn_qkv/attn_output/attn_gate, routed experts (ffn_*_exps) | Q4_K |
| shared expert (ffn_*_shexp), output.weight | Q6_K |
| norms, SSM gates | F32 |
~4.56 BPW. Container ~39.9 GB (routed experts stored raw for CPU-offload support).
Usage
1export PRISMX_PQM_STANDALONE=1
2export PRISMX_PQM=/path/to/agents-a1-malay35b-dare.pqm
3export PRISMX_TOKENIZER=/path/to/malay35b.tok
4prism_server 0.0.0.0:8080 --max-batch 1 --max-seq 4096
5# 8 GB GPU:
6prism_server 0.0.0.0:8080 --n-cpu-moe 38 --moe-cache-experts 512 --max-batch 1
Standalone mode always binds 0.0.0.0:8080. OpenAI-compatible /v1/completions and
/v1/chat/completions. Use temperature >= 0.15 (greedy decoding can collapse).
Verification
Boot-verified GGUF-free on A100 (sm_80): coherent English and Malay completions,
correct arithmetic (17×23=391), no NaN/garbage output.
Known limitations
- Text-only.
- Intermittent sub-word token drop/duplication on some code/arithmetic prompts is a
known open issue; validate structured output downstream.
