newt-core 0.7.1

Newt-Agent core types, errors, and the NeMoCode-style tier router
Documentation 
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

# Built-in model card: Ornith-1.0-35B (deepreinforce-ai) — the reference card.
# IDENTITIES + serving profile ONLY: no host / IP / GPU / port. The endpoint and
# any node-specific knob (tensor_parallel, gpu_mem) live in the operator's local
# [dgx] config or a `~/.newt/models/ornith-1.0-35b.toml` override.
name = "Ornith-1.0-35B"
backend = "vllm"
footprint_gib = 35.0 # FP8 resident footprint (the dgx_registry FP8 row).
# reasoning_parser/tool_call_parser/enable_auto_tool_choice come from the
# qwen3 family defaults (cards/families/qwen3.toml) — override here only if
# this specific card needs to diverge from the family.
family = "qwen3"

[vllm]
served_name = "Ornith-1.0-35B"
max_model_len = 262144
# Prefix caching + trust-remote-code (FP8 loads the model's own code) as raw argv
# until/unless they become typed VllmProfile fields.
extra = ["--enable-prefix-caching", "--trust-remote-code"]

[ollama]
tag = "ornith:35b"
# Capped WELL below the 262144 native window on purpose: ollama's auto num_ctx
# takes ~80% of the declared window, which for a Q4 35B (~21 GiB of weights) sizes
# a ~209K-token KV cache and OOMs a modest node. 32768 is a safe, generous default;
# override it for a bigger box.
num_ctx = 32768

[tuning]
temperature = 0.6 # the card's agentic default (1.0 reproduces the benchmarks).
top_p = 0.95
top_k = 20
context_tokens = 262144

[capability]
emits_leading_reasoning = true # qwen3.5 family: opens the turn with a <think> block.
thinking_default = true        # thinking on by default (the per-turn toggle's fail-safe).
reasoning_content_field = "reasoning_content"