defaults.yaml

# NeMoCode defaults — NVIDIA NIM full catalog
# All model specs verified against March 2026 NVIDIA releases.
# DGX Spark optimized: 128GB unified memory runs Super + Nano concurrently.

default_endpoint: nim-super
active_formation: null

# ---------------------------------------------------------------------------
# Endpoints — Nemotron family + full NIM catalog
# ---------------------------------------------------------------------------
endpoints:

  # ═══════════════════════════════════════════════════════════════════════════
  # CORE CODING MODELS (API Catalog — build.nvidia.com)
  # ═══════════════════════════════════════════════════════════════════════════

  # ── Nemotron 3 Super (primary workhorse) ──────────────────────────
  nim-super:
    name: "Nemotron 3 Super (API Catalog)"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/nemotron-3-super-120b-a12b"
    capabilities: [chat, code]
    max_tokens: 16384

  # ── Nemotron 3 Nano 30B (fast targeted subtasks) ─────────────────
  nim-nano:
    name: "Nemotron 3 Nano 30B (API Catalog)"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/nemotron-3-nano-30b-a3b"
    capabilities: [chat, code]
    max_tokens: 16384

  # ── Nemotron Nano 9B v2 (ultra-lightweight mini-agent) ───────────
  nim-nano-9b:
    name: "Nemotron Nano 9B v2 (API Catalog)"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/nemotron-nano-9b-v2"
    capabilities: [chat, code]
    max_tokens: 8192

  # ── Nemotron 3 Ultra (placeholder — not yet public) ──────────────
  # ~500B total / ~50B active. Repoint when it ships.
  nim-ultra:
    name: "Nemotron 3 Ultra (placeholder)"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/nemotron-3-ultra"
    capabilities: [chat, code]
    max_tokens: 32768

  # ═══════════════════════════════════════════════════════════════════════════
  # VISION / MULTIMODAL (API Catalog)
  # ═══════════════════════════════════════════════════════════════════════════

  # ── Nemotron Nano 12B VL (vision-language, screenshots/UI) ───────
  nim-vlm:
    name: "Nemotron Nano 12B VL v2 (API Catalog)"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/nemotron-nano-12b-v2-vl"
    capabilities: [chat, vlm]
    max_tokens: 8192

  # ── Llama Nemotron Nano VL 8B (document intelligence) ────────────
  nim-docvlm:
    name: "Llama Nemotron Nano VL 8B (API Catalog)"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/llama-3.1-nemotron-nano-vl-8b-v1"
    capabilities: [chat, vlm, parse]
    max_tokens: 8192

  # ── Cosmos Nemotron (world foundation model) ─────────────────────
  nim-cosmos:
    name: "Cosmos Nemotron (API Catalog)"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/cosmos-nemotron"
    capabilities: [vlm, image]
    max_tokens: 4096

  # ═══════════════════════════════════════════════════════════════════════════
  # REASONING / SPECIALIZED (API Catalog)
  # ═══════════════════════════════════════════════════════════════════════════

  # ── Llama Nemotron Ultra 253B (cloud-only heavyweight reasoning) ──
  nim-llama-ultra:
    name: "Llama Nemotron Ultra 253B (API Catalog)"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/llama-nemotron-ultra-253b"
    capabilities: [chat, code]
    max_tokens: 32768

  # ── Llama Nemotron Reasoning Super 49B ───────────────────────────
  nim-reasoning-super:
    name: "Llama Nemotron Reasoning Super 49B (API Catalog)"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/llama-nemotron-reasoning-super-49b"
    capabilities: [chat, code]
    max_tokens: 16384

  # ── Llama Nemotron Reasoning 8B (lightweight reasoning) ──────────
  nim-reasoning-8b:
    name: "Llama Nemotron Reasoning 8B (API Catalog)"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/llama-nemotron-reasoning-8b"
    capabilities: [chat, code]
    max_tokens: 16384

  # ═══════════════════════════════════════════════════════════════════════════
  # RETRIEVAL MODELS (API Catalog)
  # ═══════════════════════════════════════════════════════════════════════════

  nim-embed:
    name: "Nemotron Embed 1B v2"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/llama-nemotron-embed-1b-v2"
    capabilities: [embed]

  nim-rerank:
    name: "Nemotron Rerank 1B v2"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/llama-nemotron-rerank-1b-v2"
    capabilities: [rerank]

  nim-parse:
    name: "Nemotron Parse"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/nemotron-parse"
    capabilities: [parse]

  # ═══════════════════════════════════════════════════════════════════════════
  # SPEECH (API Catalog)
  # ═══════════════════════════════════════════════════════════════════════════

  # ── Parakeet ASR (speech-to-text) ────────────────────────────────
  nim-asr:
    name: "Parakeet CTC 1.1B (API Catalog)"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/parakeet-ctc-1.1b"
    capabilities: [speech]

  # ── FastPitch + HiFi-GAN TTS (text-to-speech) ───────────────────
  nim-tts:
    name: "FastPitch HiFi-GAN TTS (API Catalog)"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/fastpitch-hifigan-tts"
    capabilities: [speech]

  # ═══════════════════════════════════════════════════════════════════════════
  # SAFETY / GUARDRAILS (API Catalog)
  # ═══════════════════════════════════════════════════════════════════════════

  nim-guardrails:
    name: "NeMo Guardrails (API Catalog)"
    tier: dev-hosted
    base_url: "https://integrate.api.nvidia.com/v1"
    api_key_env: NVIDIA_API_KEY
    model_id: "nvidia/nemo-guardrails"
    capabilities: [chat]
    max_tokens: 4096

  # ═══════════════════════════════════════════════════════════════════════════
  # THIRD-PARTY HOSTED
  # ═══════════════════════════════════════════════════════════════════════════

  openrouter-super:
    name: "Nemotron 3 Super (OpenRouter)"
    tier: openrouter
    base_url: "https://openrouter.ai/api/v1"
    api_key_env: OPENROUTER_API_KEY
    model_id: "nvidia/nemotron-3-super-120b-a12b:free"
    capabilities: [chat, code]
    max_tokens: 16384
    extra_headers:
      HTTP-Referer: "https://github.com/nemocode"
      X-Title: "NeMoCode"

  together-super:
    name: "Nemotron 3 Super (Together AI)"
    tier: inference-partner
    base_url: "https://api.together.xyz/v1"
    api_key_env: TOGETHER_API_KEY
    model_id: "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8"
    capabilities: [chat, code]
    max_tokens: 16384

  deepinfra-super:
    name: "Nemotron 3 Super (DeepInfra)"
    tier: inference-partner
    base_url: "https://api.deepinfra.com/v1/openai"
    api_key_env: DEEPINFRA_API_KEY
    model_id: "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B"
    capabilities: [chat, code]
    max_tokens: 16384

  # ═══════════════════════════════════════════════════════════════════════════
  # DGX SPARK LOCAL DEPLOYMENT
  # 128GB unified memory — run Super + Nano concurrently!
  # ═══════════════════════════════════════════════════════════════════════════

  # ── Super on Spark (main brain, port 8000) ───────────────────────
  spark-nim-super:
    name: "Local NIM Super (DGX Spark)"
    tier: local-nim
    base_url: "http://localhost:8000/v1"
    model_id: "nvidia/nemotron-3-super-120b-a12b"
    capabilities: [chat, code]
    max_tokens: 16384

  # ── Nano 30B on Spark (fast worker, port 8001) ──────────────────
  spark-nim-nano:
    name: "Local NIM Nano 30B (DGX Spark)"
    tier: local-nim
    base_url: "http://localhost:8001/v1"
    model_id: "nvidia/nemotron-3-nano-30b-a3b"
    capabilities: [chat, code]
    max_tokens: 16384

  # ── Nano 9B on Spark (ultra-light mini-agent, port 8002) ────────
  spark-nim-nano9b:
    name: "Local NIM Nano 9B (DGX Spark)"
    tier: local-nim
    base_url: "http://localhost:8002/v1"
    model_id: "nvidia/nemotron-nano-9b-v2"
    capabilities: [chat, code]
    max_tokens: 8192

  # ── VLM on Spark (vision, port 8003) ────────────────────────────
  spark-nim-vlm:
    name: "Local NIM VLM 12B (DGX Spark)"
    tier: local-nim
    base_url: "http://localhost:8003/v1"
    model_id: "nvidia/nemotron-nano-12b-v2-vl"
    capabilities: [chat, vlm]
    max_tokens: 8192

  # ── Embed on Spark (retrieval, port 8004) ────────────────────────
  spark-nim-embed:
    name: "Local NIM Embed (DGX Spark)"
    tier: local-nim
    base_url: "http://localhost:8004/v1"
    model_id: "nvidia/llama-nemotron-embed-1b-v2"
    capabilities: [embed]

  # ── Rerank on Spark (retrieval, port 8005) ───────────────────────
  spark-nim-rerank:
    name: "Local NIM Rerank (DGX Spark)"
    tier: local-nim
    base_url: "http://localhost:8005/v1"
    model_id: "nvidia/llama-nemotron-rerank-1b-v2"
    capabilities: [rerank]

  # ═══════════════════════════════════════════════════════════════════════════
  # DGX SPARK vLLM DEPLOYMENT
  # Alternative to NIM containers — direct HuggingFace model serving.
  # No tensor parallel needed on Spark (128GB unified memory).
  # ═══════════════════════════════════════════════════════════════════════════

  # ── Super via vLLM on Spark (NVFP4, port 8000) ────────────────
  spark-vllm-super:
    name: "Local vLLM Super NVFP4 (DGX Spark)"
    tier: local-vllm
    base_url: "http://localhost:8000/v1"
    model_id: "/home/hmbown/HF_Models/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4"
    capabilities: [chat, code]
    max_tokens: 16384

  # ── Nano 9B via vLLM on Spark (port 8001) ──────────────────────
  spark-vllm-nano9b:
    name: "Local vLLM Nano 9B (DGX Spark)"
    tier: local-vllm
    base_url: "http://localhost:8001/v1"
    model_id: "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
    capabilities: [chat, code]
    max_tokens: 8192

  # ═══════════════════════════════════════════════════════════════════════════
  # DGX SPARK SGLANG DEPLOYMENT
  # Best local path for Nemotron 3 Super long context on Spark.
  # ═══════════════════════════════════════════════════════════════════════════

  # ── Super via SGLang on Spark (NVFP4, port 8000) ───────────────
  spark-sglang-super:
    name: "Local SGLang Super NVFP4 (DGX Spark)"
    tier: local-sglang
    base_url: "http://localhost:8000/v1"
    model_id: "nvidia/nemotron-3-super-120b-a12b"
    capabilities: [chat, code]
    max_tokens: 16384

  # ── Nano 9B via SGLang on Spark (port 8001) ────────────────────
  spark-sglang-nano9b:
    name: "Local SGLang Nano 9B (DGX Spark)"
    tier: local-sglang
    base_url: "http://localhost:8001/v1"
    model_id: "nvidia/nemotron-nano-9b-v2"
    capabilities: [chat, code]
    max_tokens: 8192

  # ═══════════════════════════════════════════════════════════════════════════
  # DGX SPARK TENSORRT-LLM DEPLOYMENT
  # Native NVIDIA TensorRT serving with OpenAI-compatible APIs.
  # ═══════════════════════════════════════════════════════════════════════════

  # ── Super via TensorRT-LLM on Spark (port 8000) ────────────────
  spark-trt-llm-super:
    name: "Local TensorRT-LLM Super 120B (DGX Spark)"
    tier: local-trt-llm
    base_url: "http://localhost:8000/v1"
    model_id: "nvidia/nemotron-3-super-120b-a12b"
    capabilities: [chat, code]
    max_tokens: 16384

  # ── Nano 4B via TensorRT-LLM on Spark (port 8001) ──────────────
  spark-trt-llm-nano4b:
    name: "Local TensorRT-LLM Nano 4B (DGX Spark)"
    tier: local-trt-llm
    base_url: "http://localhost:8001/v1"
    model_id: "nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8"
    capabilities: [chat, code]
    max_tokens: 8192

  # ═══════════════════════════════════════════════════════════════════════════
  # GENERIC LOCAL DEPLOYMENT (non-Spark workstations)
  # ═══════════════════════════════════════════════════════════════════════════

  local-nim-super:
    name: "Local NIM (Super)"
    tier: local-nim
    base_url: "http://localhost:8000/v1"
    model_id: "nvidia/nemotron-3-super-120b-a12b"
    capabilities: [chat, code]
    max_tokens: 16384

  local-nim-nano:
    name: "Local NIM (Nano)"
    tier: local-nim
    base_url: "http://localhost:8000/v1"
    model_id: "nvidia/nemotron-3-nano-30b-a3b"
    capabilities: [chat, code]
    max_tokens: 16384

  local-vllm-super:
    name: "Local vLLM (Super)"
    tier: local-vllm
    base_url: "http://localhost:8000/v1"
    model_id: "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8"
    capabilities: [chat, code]
    max_tokens: 16384

  local-vllm-nano:
    name: "Local vLLM (Nano 30B)"
    tier: local-vllm
    base_url: "http://localhost:8000/v1"
    model_id: "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8"
    capabilities: [chat, code]
    max_tokens: 16384

  local-vllm-nano9b:
    name: "Local vLLM (Nano 9B)"
    tier: local-vllm
    base_url: "http://localhost:8000/v1"
    model_id: "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
    capabilities: [chat, code]
    max_tokens: 8192

  local-sglang-super:
    name: "Local SGLang (Super)"
    tier: local-sglang
    base_url: "http://localhost:8000/v1"
    model_id: "nvidia/nemotron-3-super-120b-a12b"
    capabilities: [chat, code]
    max_tokens: 16384

  local-sglang-nano9b:
    name: "Local SGLang (Nano 9B)"
    tier: local-sglang
    base_url: "http://localhost:8000/v1"
    model_id: "nvidia/nemotron-nano-9b-v2"
    capabilities: [chat, code]
    max_tokens: 8192

  local-sglang-nano4b:
    name: "Local SGLang (Nano 4B)"
    tier: local-sglang
    base_url: "http://localhost:8000/v1"
    model_id: "nvidia/llama-3.1-nemotron-nano-4b-v1.1"
    capabilities: [chat, code]
    max_tokens: 8192

  local-trt-llm-super:
    name: "Local TensorRT-LLM (Super 120B)"
    tier: local-trt-llm
    base_url: "http://localhost:8000/v1"
    model_id: "nvidia/nemotron-3-super-120b-a12b"
    capabilities: [chat, code]
    max_tokens: 16384

  local-trt-llm-nano4b:
    name: "Local TensorRT-LLM (Nano 4B)"
    tier: local-trt-llm
    base_url: "http://localhost:8000/v1"
    model_id: "nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8"
    capabilities: [chat, code]
    max_tokens: 8192

# ---------------------------------------------------------------------------
# Manifests — model architecture specifics
# ---------------------------------------------------------------------------
manifests:

  # ═══════════════════════════════════════════════════════════════════════════
  # NEMOTRON 3 FAMILY
  # ═══════════════════════════════════════════════════════════════════════════

  nvidia/nemotron-3-super-120b-a12b:
    display_name: "Nemotron 3 Super"
    arch: nemotron-3
    tier_in_family: super
    capabilities: [chat, code]
    context_window: 1048576          # 1M tokens native
    max_output_tokens: 16384
    moe:
      total_params_b: 120
      active_params_b: 12
      uses_latent_moe: true          # 4x expert specialization
      uses_mtp: true                 # Multi-token prediction / native speculative decoding
      uses_mamba: true               # Hybrid Mamba-Transformer backbone
      precision: "nvfp4"             # Pretrained in NVFP4 on Blackwell
    reasoning:
      supports_thinking: true
      thinking_param: "enable_thinking"
      thinking_budget_param: "reasoning_budget"
      no_think_tag: "/no_think"
      supports_budget_control: true  # Granular inference-time reasoning budget
    structured:
      json_schema: true
      regex: true
      choices: true
    supports_tools: true
    supports_parallel_tools: false
    supports_lora: true
    recommended_gpu: "DGX Spark / B200 / H100"
    min_gpu_memory_gb: 80
    min_gpus: 1

  nvidia/nemotron-3-nano-30b-a3b:
    display_name: "Nemotron 3 Nano 30B"
    arch: nemotron-3
    tier_in_family: nano
    capabilities: [chat, code]
    context_window: 1048576          # 1M tokens native
    max_output_tokens: 16384
    moe:
      total_params_b: 31.6
      active_params_b: 3.2
      uses_latent_moe: false
      uses_mtp: false
      uses_mamba: true
      precision: "fp8"
    reasoning:
      supports_thinking: true
      thinking_param: "enable_thinking"
      thinking_budget_param: "reasoning_budget"
      supports_budget_control: true
    structured:
      json_schema: true
      regex: true
    supports_tools: true
    supports_lora: true
    recommended_gpu: "DGX Spark / H100 / B200"
    min_gpu_memory_gb: 24
    min_gpus: 1

  nvidia/nemotron-nano-9b-v2:
    display_name: "Nemotron Nano 9B v2"
    arch: nemotron-3
    tier_in_family: nano-9b
    capabilities: [chat, code]
    context_window: 131072           # 128K tokens
    max_output_tokens: 8192
    moe:
      total_params_b: 9
      active_params_b: 9             # Dense model, all params active
      uses_latent_moe: false
      uses_mtp: false
      uses_mamba: true               # Hybrid Mamba-Transformer
      precision: "fp8"
    reasoning:
      supports_thinking: true
      thinking_param: "enable_thinking"
      supports_budget_control: false
    structured:
      json_schema: true
      regex: true
    supports_tools: true
    supports_lora: true
    recommended_gpu: "DGX Spark / L40S / RTX 4090"
    min_gpu_memory_gb: 10
    min_gpus: 1

  nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8:
    display_name: "Nemotron 3 Nano 4B FP8"
    arch: nemotron-3
    tier_in_family: nano-4b
    capabilities: [chat, code]
    context_window: 262144
    max_output_tokens: 8192
    moe:
      total_params_b: 4
      active_params_b: 4
      precision: "fp8"
    reasoning:
      supports_thinking: true
      thinking_param: "enable_thinking"
      supports_budget_control: false
    structured:
      json_schema: true
    supports_tools: true
    recommended_gpu: "DGX Spark / GeForce RTX / Jetson Thor"
    min_gpu_memory_gb: 4
    min_gpus: 1

  nvidia/llama-3.1-nemotron-nano-4b-v1.1:
    display_name: "Nemotron Nano 4B v1.1"
    arch: other
    tier_in_family: nano-4b
    capabilities: [chat, code]
    context_window: 131072
    max_output_tokens: 8192
    moe:
      total_params_b: 4
      active_params_b: 4
      precision: "fp8"
    reasoning:
      supports_thinking: true
      thinking_param: "enable_thinking"
      supports_budget_control: false
    structured:
      json_schema: true
    supports_tools: true
    supports_lora: true
    recommended_gpu: "Any (2GB+ VRAM)"
    min_gpu_memory_gb: 2
    min_gpus: 1

  nvidia/nemotron-3-ultra:
    display_name: "Nemotron 3 Ultra (unreleased)"
    arch: nemotron-3
    tier_in_family: ultra
    capabilities: [chat, code]
    context_window: 1048576
    max_output_tokens: 32768
    moe:
      total_params_b: 500
      active_params_b: 50
      uses_latent_moe: true
      uses_mtp: true
      uses_mamba: true
      precision: "nvfp4"
    reasoning:
      supports_thinking: true
      thinking_param: "enable_thinking"
      thinking_budget_param: "reasoning_budget"
      supports_budget_control: true
    supports_tools: true
    min_gpu_memory_gb: 640
    min_gpus: 8
    recommended_gpu: "8x H100 / 8x B200"

  # ═══════════════════════════════════════════════════════════════════════════
  # VISION / MULTIMODAL
  # ═══════════════════════════════════════════════════════════════════════════

  nvidia/nemotron-nano-12b-v2-vl:
    display_name: "Nemotron Nano 12B VL v2"
    arch: nemotron-vlm
    tier_in_family: nano-vlm
    capabilities: [chat, vlm]
    context_window: 131072
    max_output_tokens: 8192
    moe:
      total_params_b: 12
      active_params_b: 12
      uses_mamba: true
      precision: "fp8"
    reasoning:
      supports_thinking: true
      thinking_param: "enable_thinking"
      supports_budget_control: false
    structured:
      json_schema: true
    supports_tools: false
    recommended_gpu: "DGX Spark / L40S / RTX 4090"
    min_gpu_memory_gb: 16
    min_gpus: 1

  nvidia/llama-3.1-nemotron-nano-vl-8b-v1:
    display_name: "Llama Nemotron Nano VL 8B"
    arch: nemotron-vlm
    tier_in_family: nano-docvlm
    capabilities: [chat, vlm, parse]
    context_window: 131072
    max_output_tokens: 8192
    moe:
      total_params_b: 8
      active_params_b: 8
      precision: "fp8"
    supports_tools: false
    recommended_gpu: "DGX Spark / L40S / RTX 4090"
    min_gpu_memory_gb: 10
    min_gpus: 1

  nvidia/cosmos-nemotron:
    display_name: "Cosmos Nemotron"
    arch: cosmos
    tier_in_family: ""
    capabilities: [vlm, image]
    context_window: 16384
    max_output_tokens: 4096
    supports_tools: false
    recommended_gpu: "H100 / B200"
    min_gpu_memory_gb: 80
    min_gpus: 1

  # ═══════════════════════════════════════════════════════════════════════════
  # REASONING / SPECIALIZED
  # ═══════════════════════════════════════════════════════════════════════════

  nvidia/llama-nemotron-ultra-253b:
    display_name: "Llama Nemotron Ultra 253B"
    arch: other
    tier_in_family: ultra
    capabilities: [chat, code]
    context_window: 131072
    max_output_tokens: 32768
    moe:
      total_params_b: 253
      active_params_b: 253           # Dense
      precision: "fp8"
    reasoning:
      supports_thinking: true
      thinking_param: "enable_thinking"
      supports_budget_control: true
    structured:
      json_schema: true
      regex: true
      choices: true
    supports_tools: true
    supports_parallel_tools: true
    recommended_gpu: "8x H100 / 8x B200"
    min_gpu_memory_gb: 640
    min_gpus: 8

  nvidia/llama-nemotron-reasoning-super-49b:
    display_name: "Llama Nemotron Reasoning Super 49B"
    arch: other
    tier_in_family: super
    capabilities: [chat, code]
    context_window: 131072
    max_output_tokens: 16384
    moe:
      total_params_b: 49
      active_params_b: 49
      precision: "fp8"
    reasoning:
      supports_thinking: true
      thinking_param: "enable_thinking"
      supports_budget_control: true
    structured:
      json_schema: true
      regex: true
    supports_tools: true
    recommended_gpu: "H100 / B200"
    min_gpu_memory_gb: 48
    min_gpus: 1

  nvidia/llama-nemotron-reasoning-8b:
    display_name: "Llama Nemotron Reasoning 8B"
    arch: other
    tier_in_family: nano
    capabilities: [chat, code]
    context_window: 131072
    max_output_tokens: 16384
    moe:
      total_params_b: 8
      active_params_b: 8
      precision: "fp8"
    reasoning:
      supports_thinking: true
      thinking_param: "enable_thinking"
      supports_budget_control: false
    structured:
      json_schema: true
    supports_tools: true
    recommended_gpu: "DGX Spark / L40S / RTX 4090"
    min_gpu_memory_gb: 10
    min_gpus: 1

  # ═══════════════════════════════════════════════════════════════════════════
  # RETRIEVAL
  # ═══════════════════════════════════════════════════════════════════════════

  nvidia/llama-nemotron-embed-1b-v2:
    display_name: "Nemotron Embed 1B v2"
    arch: nemotron-embed
    tier_in_family: ""
    capabilities: [embed]
    context_window: 32768
    supports_tools: false
    min_gpu_memory_gb: 2
    min_gpus: 1

  nvidia/llama-nemotron-rerank-1b-v2:
    display_name: "Nemotron Rerank 1B v2"
    arch: nemotron-rerank
    tier_in_family: ""
    capabilities: [rerank]
    context_window: 32768
    supports_tools: false
    min_gpu_memory_gb: 2
    min_gpus: 1

  nvidia/nemotron-parse:
    display_name: "Nemotron Parse"
    arch: nemotron-parse
    tier_in_family: ""
    capabilities: [parse]
    supports_tools: false

  # ═══════════════════════════════════════════════════════════════════════════
  # SPEECH
  # ═══════════════════════════════════════════════════════════════════════════

  nvidia/parakeet-ctc-1.1b:
    display_name: "Parakeet CTC 1.1B ASR"
    arch: speech-asr
    tier_in_family: ""
    capabilities: [speech]
    supports_tools: false
    min_gpu_memory_gb: 4
    min_gpus: 1

  nvidia/fastpitch-hifigan-tts:
    display_name: "FastPitch HiFi-GAN TTS"
    arch: speech-tts
    tier_in_family: ""
    capabilities: [speech]
    supports_tools: false
    min_gpu_memory_gb: 4
    min_gpus: 1

  # ═══════════════════════════════════════════════════════════════════════════
  # SAFETY
  # ═══════════════════════════════════════════════════════════════════════════

  nvidia/nemo-guardrails:
    display_name: "NeMo Guardrails"
    arch: guardrails
    tier_in_family: ""
    capabilities: [chat]
    context_window: 8192
    max_output_tokens: 4096
    supports_tools: false
    min_gpu_memory_gb: 4
    min_gpus: 1

# ---------------------------------------------------------------------------
# Formations — matches NVIDIA deployment patterns
# ---------------------------------------------------------------------------
formations:

  # ═══════════════════════════════════════════════════════════════════════════
  # STANDARD FORMATIONS (hosted / any hardware)
  # ═══════════════════════════════════════════════════════════════════════════

  # Default: Super does everything. The daily driver.
  solo:
    name: "Solo"
    description: "Nemotron 3 Super for day-to-day coding. Single model, full tool use."
    slots:
      - role: executor
        endpoint: nim-super
    tools: [fs, git, bash, rg]
    verification_rounds: 0

  # NVIDIA's recommended Super + Nano pattern.
  # Super handles complex reasoning/planning; Nano handles fast subtasks.
  super-nano:
    name: "Super + Nano"
    description: "Super plans and reviews. Nano executes fast targeted steps."
    slots:
      - role: planner
        endpoint: nim-super
      - role: fast
        endpoint: nim-nano
      - role: executor
        endpoint: nim-super
      - role: reviewer
        endpoint: nim-super
        reasoning_mode: "low_effort"
    tools: [fs, git, bash, rg]
    verification_rounds: 1

  # Full stack when Ultra ships.
  ultra-swarm:
    name: "Ultra Swarm"
    description: "Ultra plans, Super executes, Ultra reviews. Repoint when Ultra ships."
    slots:
      - role: planner
        endpoint: nim-ultra
      - role: executor
        endpoint: nim-super
      - role: reviewer
        endpoint: nim-ultra
    tools: [fs, git, bash, rg]
    verification_rounds: 1

  # Retrieval-augmented coding.
  retrieval:
    name: "Retrieval"
    description: "Embed + rerank + Super for RAG-augmented coding."
    slots:
      - role: embed
        endpoint: nim-embed
      - role: rerank
        endpoint: nim-rerank
      - role: executor
        endpoint: nim-super
    tools: [fs, git, bash, rg, http]

  # Document processing.
  doc-ops:
    name: "Doc Ops"
    description: "Parse documents, embed, and reason over them."
    slots:
      - role: parse
        endpoint: nim-parse
      - role: embed
        endpoint: nim-embed
      - role: executor
        endpoint: nim-super
    tools: [fs]

  # ═══════════════════════════════════════════════════════════════════════════
  # DGX SPARK FORMATIONS (all-local, air-gapped capable)
  # ═══════════════════════════════════════════════════════════════════════════

  # Spark daily driver: Super locally, Nano 9B mini-agents.
  spark:
    name: "DGX Spark"
    description: "Super on Spark + Nano 9B mini-agents. Fully local, zero cloud dependency."
    slots:
      - role: planner
        endpoint: spark-nim-super
      - role: fast
        endpoint: spark-nim-nano9b
      - role: executor
        endpoint: spark-nim-super
      - role: reviewer
        endpoint: spark-nim-super
        reasoning_mode: "low_effort"
    tools: [fs, git, bash, rg]
    verification_rounds: 1

  # Spark swarm: Super brain, multiple Nano workers.
  spark-swarm:
    name: "DGX Spark Swarm"
    description: "Super plans and reviews. Nano 30B executes, Nano 9B explores. All local."
    slots:
      - role: planner
        endpoint: spark-nim-super
      - role: fast
        endpoint: spark-nim-nano9b
      - role: executor
        endpoint: spark-nim-nano
      - role: reviewer
        endpoint: spark-nim-super
    tools: [fs, git, bash, rg]
    verification_rounds: 1

  # Spark vision: VLM for screenshots + Super for coding.
  spark-vision:
    name: "DGX Spark Vision"
    description: "VLM reads screenshots/diagrams, Super writes code. All local."
    slots:
      - role: parse
        endpoint: spark-nim-vlm
      - role: executor
        endpoint: spark-nim-super
    tools: [fs, git, bash, rg]
    verification_rounds: 0

  # Spark RAG: full local retrieval pipeline.
  spark-retrieval:
    name: "DGX Spark Retrieval"
    description: "Embed + rerank + Super, all local on Spark."
    slots:
      - role: embed
        endpoint: spark-nim-embed
      - role: rerank
        endpoint: spark-nim-rerank
      - role: executor
        endpoint: spark-nim-super
    tools: [fs, git, bash, rg]

  # Spark full-stack: everything local — vision + retrieval + coding.
  spark-full:
    name: "DGX Spark Full Stack"
    description: "VLM + embed + rerank + Super + Nano. The complete local AI workstation."
    slots:
      - role: parse
        endpoint: spark-nim-vlm
      - role: embed
        endpoint: spark-nim-embed
      - role: rerank
        endpoint: spark-nim-rerank
      - role: planner
        endpoint: spark-nim-super
      - role: fast
        endpoint: spark-nim-nano9b
      - role: executor
        endpoint: spark-nim-super
      - role: reviewer
        endpoint: spark-nim-super
        reasoning_mode: "low_effort"
    tools: [fs, git, bash, rg, http]
    verification_rounds: 1

  # Spark vLLM: Same as spark but using vLLM instead of NIM containers.
  spark-vllm:
    name: "DGX Spark (vLLM)"
    description: "Super + Nano 9B via vLLM on Spark. No Docker required."
    slots:
      - role: planner
        endpoint: spark-vllm-super
      - role: fast
        endpoint: spark-vllm-nano9b
      - role: executor
        endpoint: spark-vllm-super
      - role: reviewer
        endpoint: spark-vllm-super
        reasoning_mode: "low_effort"
    tools: [fs, git, bash, rg]
    verification_rounds: 1

  # Spark SGLang: single-backend local Super with long-context override support.
  spark-sglang:
    name: "DGX Spark (SGLang)"
    description: "Super via SGLang on Spark. Best local path for Nemotron 3 Super long context."
    slots:
      - role: planner
        endpoint: spark-sglang-super
      - role: executor
        endpoint: spark-sglang-super
      - role: reviewer
        endpoint: spark-sglang-super
        reasoning_mode: "low_effort"
    tools: [fs, git, bash, rg]
    verification_rounds: 1

  # Spark TensorRT-LLM: native TensorRT serving with Super 120B + Nano 4B.
  spark-trt-llm:
    name: "DGX Spark (TensorRT-LLM)"
    description: "Super 120B + Nano 4B via TensorRT-LLM on Spark."
    slots:
      - role: planner
        endpoint: spark-trt-llm-super
      - role: fast
        endpoint: spark-trt-llm-nano4b
      - role: executor
        endpoint: spark-trt-llm-super
      - role: reviewer
        endpoint: spark-trt-llm-super
        reasoning_mode: "low_effort"
    tools: [fs, git, bash, rg]
    verification_rounds: 1

  # Lightweight local: Nano only on Spark or workstation.
  local:
    name: "Local"
    description: "Nano on local NIM for air-gapped / offline work."
    slots:
      - role: executor
        endpoint: local-nim-nano
    tools: [fs, git, bash, rg]
    verification_rounds: 0

  # ═══════════════════════════════════════════════════════════════════════════
  # SPECIALTY FORMATIONS (hosted)
  # ═══════════════════════════════════════════════════════════════════════════

  # Vision: VLM understands screenshots, Super writes code.
  vision:
    name: "Vision"
    description: "VLM reads screenshots/UI mockups, Super writes the code."
    slots:
      - role: parse
        endpoint: nim-vlm
      - role: executor
        endpoint: nim-super
    tools: [fs, git, bash, rg]
    verification_rounds: 0

  # Deep reasoning: Llama Nemotron Ultra for hardest problems.
  deep-reason:
    name: "Deep Reasoning"
    description: "Llama Nemotron Ultra 253B for the hardest architectural problems."
    slots:
      - role: planner
        endpoint: nim-llama-ultra
      - role: executor
        endpoint: nim-super
      - role: reviewer
        endpoint: nim-llama-ultra
    tools: [fs, git, bash, rg]
    verification_rounds: 1

# ---------------------------------------------------------------------------
# Permissions
# ---------------------------------------------------------------------------
permissions:
  allow_shell: true
  allow_file_write: true
  allow_file_delete: false
  auto_approve_reads: true
  auto_approve_shell: false
  require_confirmation: [bash_exec, file_delete, git_commit]