sprappcom/oaica-35b-malay-agents-260827

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

export PRISMX_PQM_STANDALONE=1
export PRISMX_PQM=/path/to/agents-a1-malay35b-dare.pqm
export PRISMX_TOKENIZER=/path/to/malay35b.tok
prism_server 0.0.0.0:8080 --max-batch 1 --max-seq 4096
# 8 GB GPU:
prism_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.