Skip to content

Family defaults

gmlx reads each model's family from its GGUF header and starts every request from the sampling values that the family's publisher recommends. Some families also publish values for a task, which gmlx offers as intents such as @coding and @reasoning-high.

gmlx profiles                     # every family and its intents
gmlx profiles qwen3.8-27b-ud-q6   # the resolved values of one configured model

The table lists the base values of each family and what each of its intents changes. An intent that a family does not list runs with the base values. The values come from the publishers' model cards and generation configs. The default row holds generic values for any other architecture.

Some GGUFs carry their publisher's sampling values in the header, as general.sampling.* keys. Each of these replaces the family's value, and your own profiles still win over them. gmlx profiles <id> shows the values a model ends up with.

Family GGUF architectures Base values Intents
qwen3.6 qwen35, qwen35moe, qwen3next, qwen4exp temperature=1.0 top_p=0.95 top_k=20 min_p=0.0 @coding: temperature=0.6; @instruct: temperature=0.7 top_p=0.8 presence_penalty=1.5 enable_thinking=False
qwen3 qwen3, qwen3moe, qwen3vlmoe temperature=0.6 top_p=0.95 top_k=20 min_p=0.0 @instruct: temperature=0.7 top_p=0.8 enable_thinking=False
qwen2.5 qwen2, qwen2moe temperature=0.7 top_p=0.8 top_k=20 repetition_penalty=1.05 -
gemma gemma, gemma2, gemma3, gemma3n, gemma4, diffusion-gemma temperature=1.0 top_p=0.95 top_k=64 -
gpt-oss gpt-oss temperature=1.0 top_p=1.0 thinking_start_token=<|start|>assistant<|channel|>analysis<|message|> thinking_end_token=<|end|> @reasoning-high: reasoning_effort=high; @reasoning-low: reasoning_effort=low; @reasoning-medium: reasoning_effort=medium
glm glm4, glm4moe, glm-dsa, glm5-next, glm5next temperature=1.0 top_p=0.95 -
deepseek deepseek2, deepseek4 temperature=0.6 top_p=0.95 -
deepseek41 deepseek41 temperature=1.0 top_p=0.95 @reasoning-high: reasoning_effort=high; @reasoning-low: reasoning_effort=low; @reasoning-max: reasoning_effort=max
minimax minimax-m2 temperature=1.0 top_p=0.95 top_k=40 -
minimax-m3 minimax-m3 temperature=1.0 top_p=0.95 top_k=40 thinking_start_token=<mm:think> thinking_end_token=</mm:think> -
nemotron nemotron_h_moe temperature=1.0 top_p=0.95 @instruct: enable_thinking=False; @reasoning-medium: medium_effort=True
nemotron-transcribe nemotron_h_moe named (?i)nemotron-3\.5-transcribe temperature=0.6 top_p=0.95 top_k=0 thinking_budget=16384 @instruct: temperature=0.0 top_p=1.0 enable_thinking=False
hunyuan hunyuan-moe temperature=0.7 top_p=0.8 top_k=20 repetition_penalty=1.05 -
hy3 hy_v3 temperature=0.9 thinking_start_token=<think:opensource> thinking_end_token=</think:opensource> @reasoning-high: reasoning_effort=high; @reasoning-low: reasoning_effort=low
hy4 hyv4 temperature=0.9 top_p=1.0 thinking_start_token=<think:6124c78e> thinking_end_token=</think:6124c78e> @reasoning-high: reasoning_effort=high; @reasoning-low: reasoning_effort=low
kimi kimi-k3 temperature=1.0 top_p=0.95 thinking_start_token=<|open|>think<|sep|> thinking_end_token=<|close|>think<|sep|> @reasoning-high: thinking_effort=high; @reasoning-low: thinking_effort=low; @reasoning-max: thinking_effort=max
kimi-k2 deepseek2 named (?i)\bkimi temperature=1.0 top_p=0.95 -
muse muse-glimmer temperature=1.0 top_p=0.95 top_k=64 thinking_start_token=<|start|>assistant to=self<|message|> thinking_end_token=<|eom|> @reasoning-high: reasoning_strength=high; @reasoning-low: reasoning_strength=low; @reasoning-medium: reasoning_strength=medium; @reasoning-xhigh: reasoning_strength=xhigh
muse-spark muse-spark temperature=1.0 top_p=1.0 top_k=64 thinking_start_token=<|start|>assistant to=self<|message|> thinking_end_token=<|eom|> @reasoning-high: reasoning_effort=high; @reasoning-low: reasoning_effort=low; @reasoning-medium: reasoning_effort=medium; @reasoning-xhigh: reasoning_effort=xhigh
k2-horizon k2-horizon temperature=1.0 top_p=0.95 @reasoning-high: reasoning_effort=high; @reasoning-low: reasoning_effort=low; @reasoning-medium: reasoning_effort=medium
beam beam temperature=0.7 top_p=0.99 thinking_start_token=<|start|>assistant<|channel|>analysis<|message|> thinking_end_token=<|end|> @reasoning-high: reasoning_effort=high; @reasoning-low: reasoning_effort=low; @reasoning-medium: reasoning_effort=medium
flint flint temperature=1.0 top_p=0.95 top_k=20 -
llama llama, smollm3 temperature=0.6 top_p=0.9 -
mistral mistral3 temperature=0.15 -
default (anything else) temperature=0.7 top_p=0.95 @coding: temperature=0.3; @creative: temperature=1.0 min_p=0.05

The sampling values of a family, of the header's general.sampling.* keys and of your profiles change only which token the server samples. The logprobs that a request gets do not depend on them.

How a request gets its settings explains how the defaults combine with your own profiles, and models.*.family replaces one model's detected family.