Vireqo-27VL-Plus-260818

Vireqo-27VL-Plus-260818

Unofficial, experimental vision-language bundle: the unchanged Vireqo-27B-Plus-260816 ternary language model paired with Prism ML's Q8_0 Qwen3-VL vision projector.

Vireqo-27VL-Plus-260818 adds tested image input to the 7.06 GiB Vireqo ternary branch. It is a two-file llama.cpp/LM Studio bundle, not a new monolithic GGUF and not a retrained multimodal model. The language tensor payload is byte-for-byte identical to Vireqo-27B-Plus-260816; visual capability comes from the separately loaded qwen3vl_merger projector published by Prism ML for Ternary-Bonsai-27B and Bonsai-27B.

中文摘要：本版本已经在 LM Studio 中实际通过图片颜色/形状/位置、中文 OCR、表格、柱状图读取和简单图上计算测试。它复用原 Plus 三值主 GGUF，再配套一份约 600 MiB 的官方 Q8 projector。与标准 VL 相比，它保留更大的三值语言工作点，但没有进行新的视觉训练。Thinking 必须关闭。

Required files

Keep both files in the same model directory:

Logical bundle size: 8,214,579,168 bytes, approximately 7.6504 GiB. LM Studio reports 8.21 decimal GB.

The local release filename is a symbolic-link alias so the existing language weight is not duplicated. Consequently, the underlying GGUF still contains the internal general.name value Vireqo-27B-Plus-260816. The release identity comes from the bundle directory and filename. This is intentional storage deduplication, not hidden model substitution.

What provides vision

The projector has the following verified GGUF structure:

The language GGUF has qwen35.embedding_length = 5120, M-RoPE sections [11, 11, 10, 0], Qwen3-VL image/video special tokens, and a vision-aware Qwen3.8 chat template. The projector dimension and language interface match. Plus also preserves source-exact Ternary-Bonsai token embedding, output head, and recurrent SSM controls, making this projector pairing structurally conservative.

Prism ML publishes the same 629,246,880-byte object under both repositories:

• prism-ml/Ternary-Bonsai-27B-gguf, revision abbae723028d71be674e71e1a71201a6f43fab22, file Ternary-Bonsai-27B-mmproj-Q8_0.gguf;
• prism-ml/Bonsai-27B-gguf, revision f10afb355f104535e3e3e98cf7ab7795c72bd292, file Bonsai-27B-mmproj-Q8_0.gguf.

Both resolve to the exact same LFS SHA-256 shown above. The standard and Plus VL releases therefore share one physical projector in the tested local installation.

Validation

Two independent runtime paths were accepted:

1. LM Studio's bundled llama.cpp server with explicit --mmproj;
2. LM Studio's own lms load automatic projector pairing followed by the native port 1234 API.

Both paths passed the same deterministic suite:

See vision-validation.md and lmstudio-native-vision-validation.md for complete answers. The reproducible fixtures and script are in eval/.

This validates controlled synthetic images, Chinese OCR, a small table, a bar chart, spatial relations, and simple arithmetic. It is not a natural-image benchmark, MMMU score, OCR benchmark, or broad grounding claim.

Local runtime

The device used was an Apple M5 MacBook Pro with about 12 GiB reserved for the task. Results depend on runtime version, external-drive state, context, image resolution, and other unified-memory use.

LM Studio

The tested local model key is:

vireqo-27vl-plus-260818

Recommended settings:

Select Vireqo-27VL-Plus-260818, attach an image in a new Chat, ensure the blue Think control is off, and ask a direct visual question.

Thinking warning

Thinking is not accepted. With LM Studio request defaults, the model can spend the full output allowance on reasoning tokens while returning an empty visible content field—including for a text-only capital-city prompt. Setting reasoning_effort=none and enable_thinking=false, equivalent to disabling Think in the UI, restored all text and image responses.

This is the known Vireqo thinking-channel issue, not a vision-projector failure.

Local storage note

The validation setup stored the physical files on an external exFAT volume; the local release and LM Studio directories used symbolic links. No additional physical copy of the 7.59 GB language weight was created, and both VL variants shared the same projector inode.

If you use symbolic links locally, keep the external files available while loading the bundle. This Hub release contains real GGUF file contents rather than symbolic links.

Limitations

• Thinking/reasoning must remain off.
• Natural photographs, handwriting, dense documents, fine grounding, and standard multimodal benchmarks were not comprehensively tested.
• Video input is not accepted as a release capability.
• The runtime warns that difficult Qwen-VL grounding may need at least 1024 image tokens; the tested synthetic fixtures succeeded with defaults.
• Visual answers can hallucinate even when OCR and chart probes pass.
• The language model retains all limitations of Vireqo-27B-Plus-260816, including extreme ternary compression and imperfect strict instruction following.
• This is not an official Qwen or Prism ML release and not equivalent to full-precision Qwen3.8-27B.
• Do not use it for safety-critical visual interpretation, medical images, legal evidence, or financial decisions.

Reproducibility and attribution

• bundle-provenance.json records both physical sources, symlink identity, projector structure, revisions, and runtime results.
• runtime-validation.json records accepted and rejected capabilities.
• SHA256SUMS.txt verifies both required files.
• NOTICE retains Qwen and Prism ML attribution.

The bundle is provided under Apache License 2.0, consistent with the pinned upstream repositories. Retained attribution: Created using Bonsai by Prism ML. Vireqo-27VL-Plus-260818 is an independent experimental name; no affiliation or endorsement is claimed.