newt-core 0.7.1

# Built-in model card: Ornith-1.0-397B — GATED. Almost certainly exceeds a single
# node (FP8 ~400 GiB; needs multi-node tensor-parallel), so `newt dgx card setup`
# refuses it without --force via the fit pre-flight. Shipped as DATA so the profile
# exists the moment the hardware does. IDENTITIES + serving profile ONLY.
name = "Ornith-1.0-397B"
backend = "vllm"
footprint_gib = 400.0 # FP8 (the dgx_registry FP8 row).
gated = true
# reasoning_parser/tool_call_parser/enable_auto_tool_choice come from the
# qwen3 family defaults (cards/families/qwen3.toml).
family = "qwen3"

[vllm]
served_name = "Ornith-1.0-397B"
max_model_len = 262144
extra = ["--enable-prefix-caching", "--trust-remote-code"]

[tuning]
temperature = 0.6
top_p = 0.95
top_k = 20
context_tokens = 262144

[capability]
emits_leading_reasoning = true
thinking_default = true
reasoning_content_field = "reasoning_content"