Model profiles

Every model family has sampling settings its makers recommend. Action Stations keeps a table of them and applies the matching profile to every request an OpenAI-compatible server (LM Studio, vLLM, llama.cpp) or Ollama receives. A model with no profile is called exactly as before.

What a profile holds

Field	Meaning
match	Case-insensitive pattern tested against the model id; the first profile that matches is used
sampling	temperature, and top_p, top_k, min_p where the card gives them
source	The Hugging Face generation_config.json and model card the numbers come from
temperaturePolicy	How the card’s temperature combines with the prompt’s per-stage temperature (below)
reasoning	effort (low, medium, high), switch (thinking on or off) or none, and the default level
draftModel	Optional small model of the same family for speculative decoding

Family	Sampling from the card	Temperature policy	Reasoning
Qwen3 Thinking (2507)	temperature 0.6, top_p 0.95, top_k 20, min_p 0	profile	always on
Qwen3 Instruct (2507)	temperature 0.7, top_p 0.8, top_k 20, min_p 0	stage	none
Qwen3.5	temperature 1.0, top_p 0.95, top_k 20, min_p 0	profile	switch, off (measured)
Qwen3.6	temperature 1.0, top_p 0.95, top_k 20, min_p 0	profile	switch, on
Qwen3.8	temperature 1.0, top_p 0.95, top_k 20, min_p 0	profile	effort, medium
gpt-oss	temperature 1.0, top_p 1.0	stage	effort, medium
Gemma 4	temperature 1.0, top_p 0.95, top_k 64	profile	none (the card uses a token)
GLM-4.7	temperature 1.0, top_p 0.95	stage	none
Nemotron 3.5	temperature 1.0, top_p 0.95	profile	switch, on

The Qwen cards also list presence_penalty (0 to 1.5). It is left out: a presence penalty discourages repeating a token, and the JSON the stages return repeats the same keys on every row.

Profile temperature and stage temperature

Prompt files set a temperature per stage (0 for the four checking stages, up to 0.8 for creative ones). The two combine like this:

• profile: the card’s temperature is always used. For the Qwen thinking models, gpt-oss, Gemma 4 and Nemotron the card warns that greedy or near-greedy decoding loops or degrades reasoning, so a stage’s 0 would hurt the model it is meant to steady. Stages that need repeatable output still get it from the fixed seed (options.seed) and from structured output.
• stage: the stage’s temperature is used when it sets one; the card’s value is the fallback. For models whose cards allow lower temperatures (Qwen3 Instruct, GLM-4.7, whose card uses 0 for some tasks).

top_p, top_k and min_p always come from the profile.

Output cap and reasoning

A stage’s maxTokens in prompts/*.md is sized for the answer. A thinking model spends part of the same budget on reasoning, so when a profile’s reasoning.tokens is set and thinking is on, that figure (halved at low effort, doubled at high) is added to the stage cap. The request is then limited to what the model’s context window leaves after the prompt (the window comes from the provider’s model list; if it is not listed, no limit applies). If a reply still stops at the length limit, the stage asks once more with a doubled cap (at most 12,000 tokens more, and never past the context) and logs the retry.

Provider options

Set these in the provider’s options (they are not in the settings screen: they suit a benchmark or a power user, and the defaults are what most people want).

Option	Values	Effect
modelProfile	"off"	Ignore profiles: only the stage temperature is sent
reasoningEffort	off, on, low, medium, high	Overrides the profile’s reasoning default
draftModel	a model id, or "off"	Draft model for speculative decoding

How reasoning reaches the server:

• Effort models (gpt-oss, Qwen3.8): reasoning_effort on OpenAI-compatible servers, think with the level on Ollama.
• Switch models (Qwen3.5, Qwen3.6, Nemotron 3.5): chat_template_kwargs.enable_thinking on OpenAI-compatible servers, think: true | false on Ollama.

Measured defaults

Quick suite, one run each, Apple M-series with 128 GB, 32K context, seed 11. One case per task, so treat differences of a few points as noise; the standard suite should confirm these.

Model and setting	Overall	Drafting reliability	Wall time (s)
Qwen3.5-9B MLX, old settings (thinking on)	62.6	100% (2 stages failed)	1486
Qwen3.5-9B MLX, card sampling, thinking on	63.2	100%	1034
Qwen3.5-9B MLX, card sampling, thinking off	66.6	100%	1380
gpt-oss-120b MLX, old settings	54.3	58%	231
gpt-oss-120b MLX, card sampling, medium effort	41.5	58%	186
gpt-oss-120b MLX, card sampling, low effort	41.9	58%	190
gpt-oss-120b MLX, card sampling, high effort	49.8	55%	171
gpt-oss-120b MLX, stage temperature + top_p, medium	46.6	55%	171
gpt-oss-120b GGUF, stage temperature + top_p, medium	67.3	100%	1436

What this decided:

• Qwen3.5 thinks off by default. Critique scored 44 with thinking off against 32 and 15 with it on, and the old settings (greedy stages, thinking on) lost two stages to replies cut off at the length limit. This is a single case; set reasoningEffort to on to turn thinking back on.
• gpt-oss temperature stays the stage’s. Forcing the card’s 1.0 on every stage did not score higher than the old settings; the profile adds the card’s top_p.
• gpt-oss reasoning effort stays at the card’s medium. Low, medium and high scored within noise of each other and ran in the same time on MLX.
• gpt-oss-120b uses the GGUF build. The MLX build completed only 55 to 58% of drafting stages in every setting, the GGUF build all of them, so its profile has preferredBuild: gguf with the reason. Other families prefer MLX unless a profile says otherwise.
• No profile names a draft model. LM Studio refuses speculative decoding for MLX models (“not supported for batched MLX models”), which is why the first Qwen draft run failed every stage. The adapter now repeats such a request without draft_model and stops sending it. With a GGUF target (gpt-oss-120b with gpt-oss-20b as the draft) scores and wall time matched the run without a draft (67.3 in 1425 s against 67.3 in 1436 s), so no speedup was measured.

Tokens per second were not recorded by the benchmark; wall time is the speed measure.

Speculative decoding

LM Studio accepts a draft_model field. It is sent only when a draft model is named (by the profile or options.draftModel) and the server’s model list shows it installed; otherwise the request is unchanged. The draft must share the main model’s tokenizer.

MLX on Apple Silicon

On a Mac, prefer the MLX build (format safetensors in lms ls --json) unless the model’s profile says otherwise (preferredBuild; gpt-oss-120b is measured better as GGUF, see above):

lms get https://huggingface.co/mlx-community/gpt-oss-120b-MXFP4-Q4
lms get https://huggingface.co/lmstudio-community/gemma-4-26B-A4B-it-MLX-4bit

The benchmark records the profile it used in each run file (model.inference), so runs with and without a profile can be told apart. To run without it, set BENCH_PROFILE=off; to set the reasoning level, BENCH_EFFORT=low; to name a draft model, BENCH_DRAFT=<model id>.