molmo2

Molmo2-8B open-vocabulary pointing (point_to, localize, localize_3d), int8 with the vision tower in bf16

• Kind: pointer
• Protocol: openai_chat
• Default port: 8830
• VRAM: 12.7 GB
• Load time: about 180 s
• Skills that use a pointer: pointing, manipulation, scene
• Configurations that name it: default, analytic-manipulation, no-handlers, no-reactive, no-verifier, r0, r1, r2, r3, reflexes, with-task-description

Weights

harness fetch molmo2 --deploy <machine>

Directory	Source	From
<weights>/molmo2-8b	modelscope_model	allenai/Molmo2-8B

Manifest

kind = "pointer"
summary = "Molmo2-8B open-vocabulary pointing (point_to, localize, localize_3d), int8 with the vision tower in bf16"
protocol = "openai_chat"
command = ["{python}", "-u", "services/servers/serve_vlm.py", "--model", "{weights}/molmo2-8b", "--name", "Molmo2-8B",
           "--trust-remote-code", "--quantize", "8bit", "--keep-modules", "vision_backbone,lm_head",
           "--max-new-tokens", "512", "--port", "{port}"]
port = 8830
# Measured on a 3090: 12.7 GB.
vram_gb = 12.7
load_seconds = 180
client = { backend = "openai_chat", base_url = "{url}/v1", model = "Molmo2-8B", point_format = "molmo", timeout_seconds = 300 }
needs_files = ["{weights}/molmo2-8b/config.json"]

[[weights]]
dir = "molmo2-8b"
modelscope_model = "allenai/Molmo2-8B"