LBS-27B vs Muse Glimmer 30B

LBS-27B

A 27B Qwen3.8 fine-tune for GUI grounding that zooms into screenshots over a few turns before naming a point, and inclusionAI says it also helps with fine detail in large images. It needs about 96 GB at BF16 and a client that runs its crop actions. Every result so far is inclusionAI's own, with no safety results or injection guidance.

Muse Glimmer 30B

A dense 30B multimodal cousin for local agent and coding work with images, documented tool calls and adjustable reasoning. Meta's 17 GB GGUF fits a 24 GB card and runs in llama.cpp, Ollama, vLLM, SGLang and Transformers. The main risks are one tool call per turn, no documented JSON output and a usage policy that adds limits to Apache 2.0.

LBS-27B

1. Serve with vllm serve inclusionAI/LBS-27B --tensor-parallel-size 4 --max-model-len 65536, which is what the card's wrapper script runs; keep the context at 32,768 or more for multi-turn images
2. Build the system prompt with build_system_prompt from the LBS branch's zoomin_grounding_protocol.py, send the screenshot after image_path: <name>, and don't pass tools; the model answers in plain text
3. Parse <action>crop_image(image_path=..., x1=..., y1=..., x2=..., y2=...)</action>, cut the crop yourself and return it with its mapping to the original image; allow at most two crops per turn and three per episode
4. Read the final answer as [x.xx,y.yy] on a 0 to 1000 scale of the original image, and use your own executor to click
5. Pass reasoning_effort as xhigh, medium or low only, since the template raises an error for high, and don't count on thinking_budget, which the template ignores
6. Don't turn on MTP speculative decoding; the MTP weights aren't in the repository though config.json names one MTP layer

Muse Glimmer 30B

1. Start it with llama serve -hf meta-models/Muse-Glimmer-30B-GGUF --hf-file Muse-Glimmer-30B-KQuant-17GB-Q4_K_M.gguf -ngl 99 -c 131072 --host 127.0.0.1 --port 8080 --chat-template-kwargs '{"reasoning_strength":"high"}' on llama.cpp b10353 or later, or ollama run muse-glimmer on Ollama 0.32.7 or later
2. In vLLM pass --tool-call-parser muse_glimmer --reasoning-parser muse_glimmer, in SGLang muse for both, in ExecuTorch --tool-parser atem. Without them tool calls come back as plain text
3. Set reasoning per request with chat_template_kwargs.reasoning_strength (low, medium, high or xhigh). llama.cpp and ExecuTorch don't read reasoning_effort, and ExecuTorch rejects it with a 400
4. Expect one tool call per turn and return each result before the next. ExecuTorch's server rejects tool_choice="required"
5. Keep temperature 1.0, top_p 0.95 and top_k 64, don't run greedy, and never stop on <|eom|>
6. Leave max_tokens room for reasoning, since in llama.cpp a request cut off mid-thought returns empty content with finish_reason length
7. For images add Meta's 1.4 GB mmproj GGUF, which takes the build to about 22 GB with the runtime's fifth