Two behaviours worth knowing. Medium injects nothing — xhigh and low each add an instruction, medium is the template's neutral level. And effort is inert while thinking is off, since the template only computes reasoning instructions inside the enable_thinking branch.

Enable Thinking is keyed exactly enableThinking on purpose: LM Studio promotes a boolean field whose key ends in that string into its native reasoning toggle rather than drawing a second control.

One rule that holds, and one problem that doesn't

Established empirically on LM Studio 0.4.21+2. Neither is in the published docs.

Verified: base[].key must name the weights' current folder

The key is <publisher>/<folder> lowercased, where publisher is the parent directory name under the models root — it is not a manifest field. Matching is case-insensitive, so the exact-case form works too; case is not the problem it is often assumed to be.

A stale name fails silently. The wrapper still registers, still appears in the picker, and still carries its field definitions — it just binds to nothing, so no controls render and the thinking badge disappears. Renaming the weights folder breaks this every time, and it is by far the most common cause of "my custom fields stopped working".

Unsolved: the wrapper and the raw weights both get a row

LM Studio collapses the two into a single row only for weights it downloaded itself, still sitting where it put them:

Row 4 is the decisive one: the same GGUF file that folded stopped folding once it was moved. So the association is to a downloaded artifact at its original path, not to any property of the model.yaml.

These weights were never downloaded, so there is nothing to associate. No model.yaml setting fixes this. In particular, owner naming does not — a wrapper owned by local against weights published as FiditeNemini still duplicates, which killed an earlier theory that the two merely had to differ.

Publishing the weights to HuggingFace and re-downloading does not fix it — tested, no change. What still correlates perfectly is how the wrapper got into hub/models/: every one that folds is a Hub artifact LM Studio installed itself, and every hand-copied one duplicates.

The supported way to make a real artifact is lms push — and note that model.yaml belongs on the LM Studio Hub, not HuggingFace. HF hosts weights; the Hub hosts the metadata artifact, and the HF downloader ignores a model.yaml sitting in a weights repo.

lms login && lms push          # from this folder
lms get fiditenemini/unhinged-qwen3.8-27b

Untested at time of writing. If you install that way, do not re-run ./install.sh afterwards — hand-copying over a Hub-managed artifact reintroduces exactly the problem.

Until then the trade is two rows with controls, or one row without.

When the controls don't appear

Check whether the base bound:

python3 -c "import json;d=json.load(open('$HOME/.cache/lm-studio/.internal/model-index-cache.json'));[print(m['indexedModelIdentifier'],m['virtual'].get('concreteModelOptions')) for m in d['models'] if m.get('virtual') and '@' not in m['indexedModelIdentifier']]"

An empty concreteModelOptions means it did not bind — almost always a renamed weights folder.

Other silent failures seen during this work:

• Unparseable manifest.json. A smart quote (" vs “) or an em dash in name is enough; the whole wrapper is dropped with no error, and it simply vanishes from the picker.
• Hub dir disagreeing with the manifest. hub/models/<owner>/<name>/ must match manifest owner/name, and owner must be lowercase.
• Stale UI. LM Studio reindexes on a directory change, so editing files in place often changes nothing visible. Reinstall rather than trusting an in-place edit.

Context length

262,144, and that is a hard ceiling here. LM Studio routes vision models through mlx_vlm, whose Qwen3_5RotaryEmbedding accepts no scaling config and never calls initialize_rope, so YaRN and every other RoPE knob are ignored. metadataOverrides is presentation-only, so advertising more would just be false. For longer context use the GGUF sibling, where llama.cpp takes RoPE scaling as load-time settings.

Sampling

config.operation carries Qwen's thinking-mode preset (temperature 1.0, top_p 0.95, top_k 20). min_p and repeat_penalty ship unchecked — Qwen's values for both are no-ops, so enabling them would only mislead. If you turn thinking off, Qwen recommends temperature 0.7, top_p 0.80, presence_penalty 1.5; set those by hand.

Install

./install.sh

Validates against LM Studio's schema (via the sibling package's suite) then copies model.yaml and manifest.json into ~/.cache/lm-studio/hub/models/local/unhinged-qwen3.8-27b-mlx/.

Two behaviours worth knowing. Medium injects nothing — xhigh and low each add an instruction, medium is the template's neutral level. And effort is inert while thinking is off, since the template only computes reasoning instructions inside the enable_thinking branch.

Enable Thinking is keyed exactly enableThinking on purpose: LM Studio promotes a boolean field whose key ends in that string into its native reasoning toggle rather than drawing a second control.

One rule that holds, and one problem that doesn't

Established empirically on LM Studio 0.4.21+2. Neither is in the published docs.

Verified: base[].key must name the weights' current folder

The key is <publisher>/<folder> lowercased, where publisher is the parent directory name under the models root — it is not a manifest field. Matching is case-insensitive, so the exact-case form works too; case is not the problem it is often assumed to be.

A stale name fails silently. The wrapper still registers, still appears in the picker, and still carries its field definitions — it just binds to nothing, so no controls render and the thinking badge disappears. Renaming the weights folder breaks this every time, and it is by far the most common cause of "my custom fields stopped working".

Unsolved: the wrapper and the raw weights both get a row

LM Studio collapses the two into a single row only for weights it downloaded itself, still sitting where it put them:

Row 4 is the decisive one: the same GGUF file that folded stopped folding once it was moved. So the association is to a downloaded artifact at its original path, not to any property of the model.yaml.

These weights were never downloaded, so there is nothing to associate. No model.yaml setting fixes this. In particular, owner naming does not — a wrapper owned by local against weights published as FiditeNemini still duplicates, which killed an earlier theory that the two merely had to differ.

Publishing the weights to HuggingFace and re-downloading does not fix it — tested, no change. What still correlates perfectly is how the wrapper got into hub/models/: every one that folds is a Hub artifact LM Studio installed itself, and every hand-copied one duplicates.

The supported way to make a real artifact is lms push — and note that model.yaml belongs on the LM Studio Hub, not HuggingFace. HF hosts weights; the Hub hosts the metadata artifact, and the HF downloader ignores a model.yaml sitting in a weights repo.

lms login && lms push          # from this folder
lms get fiditenemini/unhinged-qwen3.8-27b

Untested at time of writing. If you install that way, do not re-run ./install.sh afterwards — hand-copying over a Hub-managed artifact reintroduces exactly the problem.

Until then the trade is two rows with controls, or one row without.

When the controls don't appear

Check whether the base bound:

python3 -c "import json;d=json.load(open('$HOME/.cache/lm-studio/.internal/model-index-cache.json'));[print(m['indexedModelIdentifier'],m['virtual'].get('concreteModelOptions')) for m in d['models'] if m.get('virtual') and '@' not in m['indexedModelIdentifier']]"

An empty concreteModelOptions means it did not bind — almost always a renamed weights folder.

Other silent failures seen during this work:

• Unparseable manifest.json. A smart quote (" vs “) or an em dash in name is enough; the whole wrapper is dropped with no error, and it simply vanishes from the picker.
• Hub dir disagreeing with the manifest. hub/models/<owner>/<name>/ must match manifest owner/name, and owner must be lowercase.
• Stale UI. LM Studio reindexes on a directory change, so editing files in place often changes nothing visible. Reinstall rather than trusting an in-place edit.

Context length

262,144, and that is a hard ceiling here. LM Studio routes vision models through mlx_vlm, whose Qwen3_5RotaryEmbedding accepts no scaling config and never calls initialize_rope, so YaRN and every other RoPE knob are ignored. metadataOverrides is presentation-only, so advertising more would just be false. For longer context use the GGUF sibling, where llama.cpp takes RoPE scaling as load-time settings.

Sampling

config.operation carries Qwen's thinking-mode preset (temperature 1.0, top_p 0.95, top_k 20). min_p and repeat_penalty ship unchecked — Qwen's values for both are no-ops, so enabling them would only mislead. If you turn thinking off, Qwen recommends temperature 0.7, top_p 0.80, presence_penalty 1.5; set those by hand.

Install

./install.sh

Validates against LM Studio's schema (via the sibling package's suite) then copies model.yaml and manifest.json into ~/.cache/lm-studio/hub/models/local/unhinged-qwen3.8-27b-mlx/.