Default Sampling Config
Recommended sampling, reasoning, compaction, and output settings for Ultramarine.
Important: use the recommended defaults
Use temperature=1 and top_p=0.95. For Chat Completions and Messages, use
top_k=-1 (disabled). Omit sampling fields to use these defaults.
Sampling defaults by API
| Setting | Chat Completions | Responses | Messages |
|---|---|---|---|
temperature | 1 | 1 | 1 |
top_p | 0.95 | 0.95 | 0.95 |
top_k | -1 (disabled) | Not accepted; omit | -1 (disabled) |
These are both the API defaults and our recommended values. For Chat
Completions and Messages, top_k=0 also disables top-k filtering. Responses
does not accept top_k; leave it out of Responses requests.
Chat Completions also accepts min_p, presence_penalty, frequency_penalty,
repetition_penalty, logit_bias, and seed. Leave these unset to start;
tune them only after testing on your tasks. See the
Chat Completions parameters
for the supported controls.
Reasoning effort
We default to max thinking (reasoning effort) and recommend it for best results.
Use the field for your API:
| API | Field | Default and recommended value |
|---|---|---|
| Chat Completions | reasoning_effort | "max" |
| Responses | reasoning.effort | "max" |
| Messages | output_config.effort | "max" |
Test high or low if you need to use fewer reasoning tokens. Ultramarine
requires reasoning; you cannot disable it.
Context compaction
The Feather Harness
compacts long agent tasks by default in the prod profile, with these defaults:
| Harness setting | Default |
|---|---|
compaction.context_window_tokens | 131072 |
compaction.trigger_tokens | 101072 |
compaction.min_turns_before_compaction | 0 |
compaction.keep_recent_turns | 2 |
Between model calls, the harness summarizes older history once the context
passes the trigger and keeps the two most recent assistant turns. Set
compaction.enabled: false to turn it off. The train profile compacts only
with compaction.enabled: true. These are harness settings, not API request
fields.
Maximum output
We recommend 65,536 tokens per request, including reasoning and answer/tool
output. Use 32,768 as a smaller option if it works well for your tasks. With
max reasoning, smaller limits can run out before the answer.
Use the field for your API:
| API | Field |
|---|---|
| Chat Completions | max_completion_tokens (max_tokens is an alias; send only one) |
| Responses | max_output_tokens |
| Messages | max_tokens (required) |
Larger limits reserve more
rate limit allowance and
funds. In Feather Harness, set
model_settings.max_tokens for normal calls and compaction.max_tokens
separately for summaries.
For prompt changes, see Prompt Tuning. For the code-execution environment, see Packages.