"thinking_token_budget": 256Ignored in this server version. It does not reserve answer tokens or cap reasoning.
FLINT · REC MLX v0.1.0
What will rec
actually use?
Some settings work. Some fail.
Others disappear before generation starts.
Local JSON review. No server call. Your request stays in this tab.
API request
JSONCopies preserve number spellings, large seeds and nested arguments. Only unsupported top-level fields are removed.
Supported fields
Review
A server test still checks context fit, schema compilation and tool execution.
Server version and review rules
These checks target the public Rec MLX server at revision 8af99ab, dated October 22, 2026. They use its top-level request allowlist, option restrictions, inherited request types and sampling rules. The standard server default uses Flint grammar.
An explicit grammar field in your JSON takes precedence over this default.
There is no reasoning-only token allowance in this inspected version. The supported output cap covers all generated tokens. Exact remaining context and the seed's context-dependent upper bound require the running model.
The checker does not send requests or execute tools. It preserves the distinction between ignored fields and rejected settings, including empty stop arrays and Python's integer/float types.
A small difference with a real effect.
"max_tokens": 15000.0A Python float. The inherited request parser requires an integer for this field.
"max_tokens": 15000A total output cap, including reasoning and the final answer or tool call.