Ollama: MiniCPM-V 4.7 fails with `Error: Failed to load CLIP model` / `unknown projector type: minicpmv4_7` (fixed in Ollama 0.40.5)

发生场景

Source(s): https://github.com/ollama/ollama/issues/18983 Status: reported working by source, not reproduced by us.

问题

ollama run openbmb/minicpm-v4.7:35b-a3b "hi" fails about 22 seconds into loading. Text-only prompts fail too, because the vision projector loads with the model. Re-pulling and verifying checksums doesn't help. Also reproduced on macOS with the 1B variant imported from the official GGUF.

报错原文（已脱敏/规范化）

Error: Failed to load CLIP model

clip_init: failed to load model '...\blobs\sha256-<hash>': load_hparams: unknown projector type: minicpmv4_7

适用环境

Ollama 0.40.3 and 0.40.4, Windows 11 (RTX 5080) and macOS Apple Silicon.

如何确认是这个问题

1. ollama --version → expect: 0.40.4 or older.
2. server.log contains unknown projector type: minicpmv4_7.

不是这个问题时 / 边界

Failed to load CLIP model for other vision models on current Ollama (could be a corrupt download or a different unsupported projector; check the unknown projector type line in server.log).

根因

The llama.cpp version bundled in Ollama 0.40.3/0.40.4 (b11521) predates upstream support for the MiniCPM-V 4.7 projector, which landed in llama.cpp b11551 (a collaborator pointed to the upstream PR).

修复步骤

1. Update Ollama to 0.40.5 or later. → expect: the original poster confirmed the model loads and answers an image prompt correctly on 0.40.5.
2. Run the model again (our inference: the original poster only updated and did not re-pull): ollama run openbmb/minicpm-v4.7:35b-a3b.

如何验证

Send an image prompt (for example a solid red image) → expect: a correct description instead of Failed to load CLIP model.

试过没用的办法

Re-pulling the model on 0.40.4 did not help; only the update did (original poster).

风险提示

none known.

验证范围

Ollama 0.40.5, October 2026.

何时失效

n/a (older versions only).