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

Oaica 35B Malay Vision 260827

Formerly: sprappcom/katcoder-vision-pqm

Single-file multimodal .pqm — a GGUF-free, Python-free, one-file package that carries both the language model weights and the Qwen vision tower embedded as a .pqm extent.

• kat_malay35b_vision.pqm (40.76 GB) — the model: katcoder-dare (kat-coder ⊕ malay35B DARE merge) with the Qwen vision tower EMBEDDED as a native .pqm extent.
• kat_malay35b_dare_v2.tok (8.94 MB) — the matching tokenizer.

Everything a run needs — LM weights, vision encoder weights, and packed Marlin extents — lives inside the single .pqm file. No GGUF, no Python, no sidecar tower.

What this is

The base is katcoder-dare, a DARE merge of:

• Qwen3.6-35B-A3B (Alibaba, Apache-2.0) — MoE backbone (A3B active)
• KAT-Coder-V2.5-Dev (Kwaipilot, Apache-2.0) — code / tool-calling capability

with the Qwen vision tower (Alibaba, Apache-2.0) packed into the same .pqm as an embedded extent, so the file is a complete standalone multimodal model.

IMPORTANT — engine requirement

This .pqm needs the prism-engine binary built with the candle-vision in-process encoder (feature-gated). The standard shipped prism-engine binary does NOT include the vision encoder — booting this file with a stock binary will load the LM but cannot run the image path.

• The merged unified master (commit 68a605e) carries the candle-vision feature.
• Build that commit with the vision feature enabled to get an in-process candle-cuda bf16 vision encoder — no external Python vision service.

Boot

PRISMX_PQM_STANDALONE=1 prism-engine \
  --model kat_malay35b_vision.pqm \
  --tokenizer kat_malay35b_dare_v2.tok \
  --n-cpu-moe 28 \
  --moe-cache-experts 512      # use 2048 if prefill panics

• PRISMX_PQM_STANDALONE=1 — boot straight from the .pqm, GGUF-free.
• --n-cpu-moe 28 — offload 28 MoE layers' experts to CPU.
• --moe-cache-experts 512 — expert cache size; raise to 2048 if prefill panics.
• The vision path runs in-process as candle-cuda bf16 (no sidecar).

Image API

OpenAI-compatible image_url content array:

{
  "model": "katcoder-vision",
  "messages": [
    {
      "role": "user",
      "content": [
        { "type": "text", "text": "What is in this image?" },
        { "type": "image_url", "image_url": { "url": "data:image/png;base64,<...>" } }
      ]
    }
  ]
}

Measured performance

• TTFT: ~1.25–1.35 s per image
• VRAM: ~16.5 GB

License & attribution

Proprietary packaging © BCZ Singapore Pte Ltd.

The packaged model weights are built entirely from Apache-2.0 components:

Component	Source	License
Qwen3.6-35B-A3B (MoE backbone)	Alibaba	Apache-2.0
KAT-Coder-V2.5-Dev (code / tools)	Kwaipilot	Apache-2.0
Qwen vision tower (embedded extent)	Alibaba	Apache-2.0

See the bundled NOTICE file for the full Apache-2.0 text and per-component attributions.