From 56b233afb6c3e40c6328633734c3164cc3edce69 Mon Sep 17 00:00:00 2001 From: "K. Isom" Date: Wed, 16 Sep 2026 06:03:51 -0700 Subject: [PATCH] straylight: add Qwen3-VL-8B abliterated as the small VLM Ornith 1.5 9B is a dense text model, not vision. Replace that role with prithivMLmods Qwen3-VL-8B Instruct abliterated v2 (mradermacher Q4_K_M) and pin mmproj-url so router mode actually loads the projector. --- hw/straylight/default.nix | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/hw/straylight/default.nix b/hw/straylight/default.nix index 3e8c3ab..1376ec7 100644 --- a/hw/straylight/default.nix +++ b/hw/straylight/default.nix @@ -50,12 +50,20 @@ let hf-repo = deepreinforce-ai/Ornith-1.0-35B-GGUF:Q4_K_M dedup-cache-models = true - ; Ornith 1.5 9B, refusal-ablated (junafinity / mradermacher GGUF). Multimodal: - ; llama-server pulls mmproj with --hf-repo. Q4_K_M is the recommended quant. + ; Ornith 1.5 9B, refusal-ablated (junafinity / mradermacher GGUF). Dense + ; text/coding model — not a VLM; the shipped mmproj does not add vision. [ornith-1.5-9b-uncensored] hf-repo = mradermacher/Ornith-1.5-9B-uncensored-GGUF:Q4_K_M dedup-cache-models = true + ; Qwen3-VL-8B Instruct, refusal-ablated (prithivMLmods v2 / mradermacher). + ; Dedicated VLM in the same size class as Ornith 1.5 9B. Explicit mmproj-url + ; because router-mode --hf-repo does not attach the projector on its own. + [qwen3-vl-8b-abliterated] + hf-repo = mradermacher/Qwen3-VL-8B-Instruct-abliterated-v2.0-GGUF:Q4_K_M + mmproj-url = https://huggingface.co/mradermacher/Qwen3-VL-8B-Instruct-abliterated-v2.0-GGUF/resolve/main/Qwen3-VL-8B-Instruct-abliterated-v2.0.mmproj-f16.gguf + dedup-cache-models = true + ; Ornith 1.5 35B-A3B MoE (official GGUF). ~3B active, multimodal; mmproj ; comes with --hf-repo. Q4_K_M is ~22 GB. [ornith-1.5-35b-a3b]