Router Tiers¶

The classifier chooses exactly one of these six tiers for every auto-routed request.

Tier	Best for	NVIDIA example models
`LITE`	simple questions, casual chat, rewriting, extraction, formatting, short answers	`google/gemma-2-2b-it`, `google/gemma-3n-e2b-it`, `nvidia/nemotron-mini-4b-instruct`, `stepfun-ai/step-3.5-flash`
`NORMAL`	daily assistant tasks, normal coding help, summaries, ordinary reasoning	`meta/llama-3.1-8b-instruct`, `nvidia/nvidia-nemotron-nano-9b-v2`, `openai/gpt-oss-20b`, `minimaxai/minimax-m2.7`
`NORMAL_CALCULATION`	math, physics, engineering, finance, statistics, algorithms, standard quantitative reasoning	`nvidia/nemotron-3-nano-omni-30b-a3b-reasoning`, `qwen/qwen3.5-122b-a10b`, `openai/gpt-oss-120b`, `z-ai/glm-5.1`
`MEDIUM`	nuanced writing, multi-step analysis, medium-to-heavy coding, planning, stronger judgment	`meta/llama-3.3-70b-instruct`, `meta/llama-4-maverick-17b-128e-instruct`, `mistralai/mistral-medium-3.5-128b`, `moonshotai/kimi-k2.6`
`HEAVY_CALCULATION`	advanced proofs, graduate physics, long derivations, symbolic reasoning, difficult optimization	`nvidia/nemotron-3-ultra-550b-a55b`, `mistralai/mistral-large-3-675b-instruct-2512`, `nvidia/nemotron-3-nano-omni-30b-a3b-reasoning`, `openai/gpt-oss-120b`
`HEAVY`	extremely hard non-calculation reasoning, severe debugging, large architecture, high-stakes synthesis	`nvidia/nemotron-3-ultra-550b-a55b`, `mistralai/mistral-large-3-675b-instruct-2512`, `meta/llama-3.3-70b-instruct`, `meta/llama-4-maverick-17b-128e-instruct`

`LITE`¶

For: simple questions, casual chat, rewriting, extraction, formatting, short answers

NVIDIA examples in this pool:

google/gemma-2-2b-it
google/gemma-3n-e2b-it
nvidia/nemotron-mini-4b-instruct
stepfun-ai/step-3.5-flash

For: daily assistant tasks, normal coding help, summaries, ordinary reasoning

NVIDIA examples in this pool:

meta/llama-3.1-8b-instruct
nvidia/nvidia-nemotron-nano-9b-v2
openai/gpt-oss-20b
minimaxai/minimax-m2.7

For: math, physics, engineering, finance, statistics, algorithms, standard quantitative reasoning

NVIDIA examples in this pool:

nvidia/nemotron-3-nano-omni-30b-a3b-reasoning
qwen/qwen3.5-122b-a10b
openai/gpt-oss-120b
z-ai/glm-5.1

For: nuanced writing, multi-step analysis, medium-to-heavy coding, planning, stronger judgment

NVIDIA examples in this pool:

meta/llama-3.3-70b-instruct
meta/llama-4-maverick-17b-128e-instruct
mistralai/mistral-medium-3.5-128b
moonshotai/kimi-k2.6

For: advanced proofs, graduate physics, long derivations, symbolic reasoning, difficult optimization

NVIDIA examples in this pool:

nvidia/nemotron-3-ultra-550b-a55b
mistralai/mistral-large-3-675b-instruct-2512
nvidia/nemotron-3-nano-omni-30b-a3b-reasoning
openai/gpt-oss-120b

For: extremely hard non-calculation reasoning, severe debugging, large architecture, high-stakes synthesis

NVIDIA examples in this pool:

nvidia/nemotron-3-ultra-550b-a55b
mistralai/mistral-large-3-675b-instruct-2512
meta/llama-3.3-70b-instruct
meta/llama-4-maverick-17b-128e-instruct