oaica-35b-a3b-vision

oaica/oaica-35b-a3b-vision

oaica-35b-a3b-vision is a machine learning model from oaica. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.

Standalone Kwaipilot/KAT-Coder-V2.5-Dev (Qwen3.6-35B-A3B base, unmodified upstream lineage, Apache-2.0) in the proprietary .oqm container, with the Qwen3.6-35B-A3B vision tower embedded in-file as extent global.vision…

Downloads · 30 days

Access

Public

Updated Oct 7, 2026

Repo size

40.8 GB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.oqm40.8 GB · 100%

At a glance

License: other
Access: Public
Created: Aug 27, 2026
Updated: Oct 7, 2026
SHA: 49d3f091

Base models

Kwaipilot/KAT-Coder-V2.5-Dev

Model card

From the Hugging Face model README

Oaica 35B A3B Vision (260827)

Standalone Kwaipilot/KAT-Coder-V2.5-Dev (Qwen3.6-35B-A3B base, unmodified upstream lineage, Apache-2.0) in the proprietary .oqm container, with the Qwen3.6-35B-A3B vision tower embedded in-file as extent global.vision.tower.safetensors (~893 MB, 333 tensors, BF16). Single artifact + tokenizer = complete multimodal server: GGUF-free boot, in-process vision encoder, no Python bridge.

Both quant lanes carry the same embedded tower:

file	variant	size	notes
`oaica-35b-a3b-vision-260827-k4h.oqm`	k4h	40,763,801,600 B	`--cpu-moe-passthrough` k4h packing; runs on any VRAM including 0
`oaica-35b-a3b-vision-260827.pqm`	byte-identical legacy name	40,763,801,600 B	same sha256; kept because the id shipped
`oaica-35b-a3b-vision-260827-bf16.oqm`	bf16	70,260,285,440 B	full-precision; A100-class VRAM or `--gpu-mem <GiB>` offload
`oaica-35b-a3b-vision-260827.tok`	tokenizer		`apexneu_unified_48k_v5`

Serve

export OAICA_OQM_STANDALONE=1
export OAICA_OQM=<path>/oaica-35b-a3b-vision-260827-k4h.oqm
export OAICA_TOKENIZER=<path>/oaica-35b-a3b-vision-260827.tok
./oaica_server "$OAICA_OQM" 0.0.0.0:<port> \
  --n-cpu-moe 28 --moe-cache-experts 2048 --max-batch 1 --max-seq 8192

OpenAI-shape image_url content arrays on /v1/chat/completions. TTFT with an image ~1.3 s. The text (code) path is unaffected by the embedded tower.

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

A transformers-loadable bf16 build (per-expert MoE layout, tower as a sidecar) is at sprappcom/oaica-35b-a3b-vision-bf16.

Siblings

model	repo
text-only	`oaica-35b-a3b` (bf16) · `oaica-35b-a3b-dev` (k4h)
Malay + vision	`oaica-35b-a3b-malay-vision`

License & attribution

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 and the verbatim Apache-2.0 text.

License: other
Languages: en, code
Created: Aug 27, 2026
Updated: Oct 7, 2026