1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
# ---------- Meta Model API (OpenAI-compatible /v1) ----------------------------
# Muse Spark is served over Meta's own endpoint as an OpenAI chat-completions
# drop-in, so it carries no dialect of its own.
#
# Every capability below was read off a live 2026-09-02 probe, not inferred
# from the family or from Meta's docs:
# - native tools: a request with a tool schema came back with
# finish_reason "tool_calls" and an OpenAI-shaped tool_calls array.
# - prompt caching: a repeated prefix reported
# prompt_tokens_details.cached_tokens > 0.
# - reasoning is native and NOT optional. usage always reports
# reasoning_tokens, and the model spends them before it emits anything:
# "what is 2+2" answered only after 274 reasoning tokens, and a 16-token
# cap returned finish_reason "length" with null content.
#
# Reasoning is declared `enabled` rather than as an effort ladder.
# `reasoning_effort` is accepted on the wire but appears inert: "low" and
# "high" spent an identical 197 reasoning tokens on the same trivial prompt.
# No effort levels are declared here, because a lever that does not move is
# worse than an absent one. Re-probe before adding them.
#
# The `-contributor` ids are the same weights on Meta's training tier, so they
# share this rule. What differs between the tiers is price and the data grant,
# which lives on the model rows' `data_controls`, not here.
[[]]
= "muse-spark-*"
= true
= "native"
= "/chat/completions"
= true
= "native"
= "native_json"
= ["enabled"]
= false
= true
= true
= true
= false
= true
= false
= false
= false
= "inline"