Decision Models¶

A judgment needs a decision model: a model built to score typed questions rather than generate text. Open-source ones are published as GGUF files on Hugging Face, and LM Studio runs the GGUF builds.

Models known to work¶

These three were verified end to end with this gem, over both decision endpoints:

Model	Hugging Face	Params	Size	Images	`decision_type`	License
D1 Omni 600M	LiquidAI/d1-omni-600M-GGUF	600M	0.8 GB	yes	`lfm2-d1-omni`	Liquid AI's own license
Kev 0.8B	ggml-org/Kev-0.8B-GGUF	753M	0.8 GB	no	`kev`	Apache-2.0
Clef Flash	ggml-org/Clef-Flash-GGUF	9.1B	7.4 GB	yes	`clef`	Apache-2.0

Download with lms get and the repository URL:

lms get https://huggingface.co/LiquidAI/d1-omni-600M-GGUF
lms get https://huggingface.co/ggml-org/Kev-0.8B-GGUF
lms get https://huggingface.co/ggml-org/Clef-Flash-GGUF

lms ls then lists them under their own DECISION heading.

Other repositories tagged decision-model on Hugging Face — the rest of the ggml-org Kev, Clef and vllm-d3 families, Laya, Julia-1, LiquidAI's d1-3B — should work too, as long as LM Studio lists them as decision models. Check each license: some (OpenJev, Bespoke-Nimble) are non-commercial.

Which one to use¶

Measured on an M2 Max with the questions from example 18_decision_models.rb (one ticket, three questions), warm model:

Model	Time per judgment	Quality, in practice
D1 Omni 600M	~30 ms	Strong on clear-cut text and spam; noticeably sensitive to wording; reads images reliably only over `:system_one`
Kev 0.8B	~65 ms	Least sure of itself; sent "I was charged twice. Please fix it today." to `other` rather than `billing`, and could not spot a gift-card scam
Clef Flash	~600 ms	Most accurate and most consistent across wording and endpoints; reads images over both endpoints

A reasonable default: start with D1 Omni 600M, tune your questions on it, and move to Clef Flash where the small model's answers aren't trustworthy enough. Examples 16-20 default to D1 Omni 600M (LMS_DECISION_MODEL overrides it).

It has to be a decision model to LM Studio¶

LM Studio decides what counts, and it must say so itself. GET /api/v1/models reports a recognized model with type: "decision":

{ "key": "d1-omni-600m", "type": "decision", "decision_type": "lfm2-d1-omni",
  "has_vision_adapter": true, "architecture": "lfm2", "format": "gguf" }

A model that merely shares a decision model's name is not enough. An MLX conversion of Clef Flash (mlx-community/clef-flash) loads as an ordinary LLM, and both decision endpoints reject it. Use the GGUF repositories above.

A chat model is refused with a RubyLLM::BadRequestError:

Invalid model identifier "qwen/qwen3.8-27b". Please specify a valid downloaded
model (e.g., d1-omni-600m, kev-0.8b, clef-flash).

With no decision model downloaded at all, the message is "There are no downloaded decision models."

How the gem lists them¶

provider.list_models reports decision models as judgment models, so you can find them programmatically:

provider = RubyLLM::Provider.resolve!(:lms).new(RubyLLM.config)
deciders = provider.list_models.select { |model| model.capabilities.include?('judgment') }

deciders.each do |model|
  puts [model.id, model.metadata[:decision_type], model.modalities.input.inspect].join('  ')
end
# clef-flash    clef          ["text", "image"]
# d1-omni-600m  lfm2-d1-omni  ["text", "image"]
# kev-0.8b      kev           ["text"]

Field	For a decision model
`capabilities`	`["judgment"]` and nothing else — they don't chat
`modalities.output`	`["judgment"]`
`modalities.input`	`["text"]`, plus `"image"` when LM Studio reports a vision adapter
`metadata[:decision_type]`	the kind of decision model, as LM Studio reports it

Loading and memory¶

Decision models load on demand like any other model, with the same idle TTL. One quirk worth knowing: if you load a model yourself (in the app, or with lms load) and a decision request then loads it on demand, LM Studio can start a second instance beside yours. lms ps shows both (clef-flash and clef-flash:2), each holding its own memory, and judgment.model reports the instance that answered — clef-flash:2, not the model id. Unload the copy you don't need with lms unload.