sprappcom/oaica-35b-a3b

Oaica 35B A3B — engine containers (260827)

Kwaipilot/KAT-Coder-V2.5-Dev — clean upstream code / tool-calling model, Qwen3.6-35B-A3B MoE, Apache-2.0, unmodified lineage (no fine-tune, no merge) — packed into the proprietary .oqm container for the oaica inference engine. GGUF-free standalone boot.

Two quant lanes, same weights, different footprints:

file	variant	size	notes
oaica-35b-a3b-260827-bf16.oqm	bf16, default	69,367,107,584 B	full-precision. 64.6 GiB resident on an A100-class GPU (--gpu-mem auto), or offload with --gpu-mem <GiB>
kat_v25_bf16.pqm	byte-identical legacy name of the above	69,367,107,584 B	same sha256; kept because the id shipped
kat_v25_raw.pqm	k4h (Q4_K/Q6_K mix)	19,737,964,544 B	~18.4 GiB, runs on any VRAM including 0 (--gpu-mem 0)
kat_v25.tok	tokenizer	8,944,940 B	apexneu_unified_48k_v5

Context: 256k native (max_position_embeddings), no YaRN or env change needed. For 1M, add OAICA_ROPE_YARN=1 --max-seq 1048576 OAICA_KV_F16=1 — a served-config change, not a different file.

Serve

export OAICA_OQM_STANDALONE=1
export OAICA_OQM=<path>/oaica-35b-a3b-260827-bf16.oqm     # or kat_v25_raw.pqm for k4h
export OAICA_TOKENIZER=<path>/kat_v25.tok
./oaica_server "$OAICA_OQM" 0.0.0.0:<port> --max-seq 262144

OAICA_OQM, OAICA_OQM_STANDALONE and OAICA_TOKENIZER are the canonical names; the older PRISMX_PQM* / PRISM_PQM* / PRISMX_TOKENIZER spellings still work — the engine rewrites them at startup and prints a deprecation line.

bf16 safetensors

For a transformers-loadable bf16 build with the vision tower, see sprappcom/oaica-35b-a3b-vision-bf16.

Siblings

model	repo
+ vision tower	oaica-35b-a3b-vision
Malay CPT/SFT	oaica-35b-a3b-malay
Malay + coder merge	oaica-35b-a3b-malay-coder
Malay + agents merge	oaica-35b-a3b-malay-agents
Malay + vision	oaica-35b-a3b-malay-vision

License & attribution

• Base model weights — Apache-2.0. © 2026 Kwaipilot / Kuaishou. These are the open Apache-2.0 KAT-Coder-V2.5-Dev weights (themselves Qwen3.6-35B-A3B-based), repacked only — not the separate closed API-only "KAT-Coder" product.
• Proprietary layer — © 2026 BCZ Singapore Pte Ltd. All rights reserved. The .oqm container format and packaging, and the oaica inference engine (Rust/CUDA), are proprietary and are not licensed under Apache-2.0.

See NOTICE for the full per-component attribution.