harn-vm 0.10.153

# NVIDIA NIM hosted LLM routes. The API Catalog exposes 100+ OpenAI-compatible
# model IDs; Harn catalogues the current agent/coding routes that are useful as
# first-class selectors and lets the live refresh workflow surface the rest.

[models."nvidia/nemotron-3-ultra-550b-a55b"]
name = "Nemotron 3 Ultra 550B A55B (NVIDIA NIM)"
provider = "nvidia"
context_window = 1048576
logical_model = "nemotron-3-ultra-550b-a55b"
equivalence_group = "nemotron-3-ultra-550b-a55b"
served_variant = "nvidia-nim"
api_dialect = "openai_chat"
capabilities = ["tools", "streaming", "thinking"]
architecture = { parameter_count_b = 550.0, active_parameter_count_b = 55.0, moe = true, source_url = "https://build.nvidia.com/nvidia/nemotron-3-ultra-550b-a55b/modelcard", last_verified = "2026-06-20" }
tier = "frontier"
open_weight = true
strengths = ["agentic", "reasoning", "coding", "tool_use", "long_context"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "deepseek", "kimi"]

[models."nvidia/nemotron-3-super-120b-a12b"]
name = "Nemotron 3 Super 120B A12B (NVIDIA NIM)"
provider = "nvidia"
context_window = 1048576
logical_model = "nemotron-3-super-120b-a12b"
equivalence_group = "nemotron-3-super-120b-a12b"
served_variant = "nvidia-nim"
api_dialect = "openai_chat"
capabilities = ["tools", "streaming", "thinking"]
architecture = { parameter_count_b = 120.0, active_parameter_count_b = 12.0, moe = true, source_url = "https://build.nvidia.com/nvidia/nemotron-3-super-120b-a12b/modelcard", last_verified = "2026-06-20" }
tier = "frontier"
open_weight = true
strengths = ["agentic", "reasoning", "coding", "tool_use", "long_context", "speed"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "deepseek", "kimi"]

# Renamed 2026-09-20: NIM now serves this as `nvidia/nemotron-nano-3-30b-a3b`.
# The old `nvidia/nemotron-3-nano-30b-a3b` id is absent from
# https://integrate.api.nvidia.com/v1/models (observed 2026-09-20); the
# `nvidia-nemotron-nano` alias follows the new id. The sibling
# `nemotron-3-nano-omni-30b-a3b-reasoning` route kept its old-style id.
[models."nvidia/nemotron-nano-3-30b-a3b"]
name = "Nemotron Nano 3 30B A3B (NVIDIA NIM)"
provider = "nvidia"
context_window = 262144
logical_model = "nemotron-nano-3-30b-a3b"
equivalence_group = "nemotron-nano-3-30b-a3b"
served_variant = "nvidia-nim"
api_dialect = "openai_chat"
capabilities = ["tools", "streaming", "thinking"]
architecture = { parameter_count_b = 30.0, active_parameter_count_b = 3.0, moe = true, source_url = "https://build.nvidia.com/nvidia/nemotron-nano-3-30b-a3b/modelcard", last_verified = "2026-09-20" }
tier = "mid"
open_weight = true
strengths = ["speed", "cheap", "reasoning", "coding", "tool_use", "long_context"]

[models."nvidia/nemotron-3.5-lightning-30b-a3b"]
name = "Nemotron 3.5 Lightning 30B A3B (NVIDIA NIM)"
provider = "nvidia"
context_window = 1048576
logical_model = "nemotron-3.5-lightning-30b-a3b"
equivalence_group = "nemotron-3.5-lightning-30b-a3b"
served_variant = "nvidia-nim"
api_dialect = "openai_chat"
capabilities = ["tools", "streaming", "thinking", "structured_output"]
architecture = { parameter_count_b = 30.0, active_parameter_count_b = 3.0, moe = true, source_url = "https://build.nvidia.com/nvidia/nemotron-3.5-lightning-30b-a3b/modelcard", last_verified = "2026-08-26" }
tier = "mid"
open_weight = true
strengths = ["speed", "coding", "agentic", "reasoning", "tool_use", "long_context"]

[models."nvidia/nemotron-3-nano-omni-30b-a3b-reasoning"]
name = "Nemotron 3 Nano Omni 30B A3B Reasoning (NVIDIA NIM)"
provider = "nvidia"
context_window = 262144
logical_model = "nemotron-3-nano-omni-30b-a3b-reasoning"
equivalence_group = "nemotron-3-nano-omni-30b-a3b-reasoning"
served_variant = "nvidia-nim"
api_dialect = "openai_chat"
capabilities = ["tools", "vision", "video", "audio", "streaming", "thinking"]
architecture = { parameter_count_b = 30.0, active_parameter_count_b = 3.0, moe = true, source_url = "https://build.nvidia.com/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning/modelcard", last_verified = "2026-06-20" }
tier = "mid"
open_weight = true
strengths = ["vision", "reasoning", "tool_use", "summarization", "long_context"]

# 2026-08-15 reachability sweep: NIM retired both undated DeepSeek V4 builds
# (HTTP 410, end of life 2026-08-07). There is no DeepSeek V4 Pro build on NIM
# any more, so the Pro row is gone rather than left pointing at a 410 — a
# catalog entry that cannot be called is worse than an absent one, because it is
# selectable. By 2026-09-26 NIM had dropped the dated V4 Flash snapshot too and
# serves DeepSeek V4.1 Flash in its place.

[models."nvidia/deepseek-v4.1-flash"]
name = "DeepSeek V4.1 Flash (NVIDIA NIM)"
provider = "nvidia"
wire_model = "deepseek-ai/deepseek-v4.1-flash"
context_window = 1048576
logical_model = "deepseek-v4.1-flash"
equivalence_group = "deepseek-v4.1-flash"
served_variant = "nvidia-nim"
api_dialect = "openai_chat"
# A one-tool `tool_choice: auto` probe on 2026-09-26 returned a native tool call
# and reported cached prompt tokens.
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
tier = "mid"
open_weight = true
strengths = ["speed", "cheap", "tool_use", "reasoning", "long_context"]

[models."nvidia/deepseek-v4-flash-0731"]
name = "DeepSeek V4 Flash 0731 (NVIDIA NIM)"
provider = "nvidia"
wire_model = "deepseek-ai/deepseek-v4-flash-0731"
context_window = 1048576
logical_model = "deepseek-v4-flash-0731"
equivalence_group = "deepseek-v4-flash"
served_variant = "nvidia-nim"
api_dialect = "openai_chat"
capabilities = ["tools", "streaming", "thinking", "prompt_caching"]
deprecated = true
deprecation_note = "Absent from NIM's model index (https://integrate.api.nvidia.com/v1/models, read 2026-09-26). NIM serves DeepSeek V4.1 Flash instead."
superseded_by = "nvidia/deepseek-v4.1-flash"
tier = "mid"
open_weight = true
strengths = ["speed", "coding", "agentic", "tool_use", "reasoning", "long_context"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "kimi"]

# Removed 2026-09-20: nvidia/minimax-m3. NIM no longer serves any MiniMax
# build — `minimaxai/minimax-m3` is absent from
# https://integrate.api.nvidia.com/v1/models (observed 2026-09-20). The
# `nvidia-minimax-m3` alias went with it; MiniMax M3 stays reachable through
# the direct MiniMax, SambaNova, Fireworks and OpenRouter rows.
[models."nvidia/kimi-k2.6"]
name = "Kimi K2.6 (NVIDIA NIM)"
provider = "nvidia"
wire_model = "moonshotai/kimi-k2.6"
context_window = 262144
logical_model = "moonshot-kimi-k2.6"
equivalence_group = "moonshot-kimi-k2.6"
served_variant = "nvidia-nim"
api_dialect = "openai_chat"
capabilities = ["tools", "vision", "streaming", "thinking", "prompt_caching"]
pricing = { input_per_mtok = 0.60, output_per_mtok = 2.50, cache_read_per_mtok = 0.15 }
tier = "frontier"
open_weight = true
strengths = ["coding", "agentic", "long_context", "tool_use", "reasoning", "vision"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "deepseek"]

# Removed 2026-09-20: nvidia/openai/gpt-oss-120b. NIM's index now carries only
# the 20B sibling below (https://integrate.api.nvidia.com/v1/models, observed
# 2026-09-20). GPT-OSS 120B remains catalogued on Cerebras, Groq, Fireworks,
# DeepInfra, SambaNova and Baseten.
[models."nvidia/openai/gpt-oss-20b"]
name = "GPT-OSS 20B (NVIDIA NIM)"
provider = "nvidia"
wire_model = "openai/gpt-oss-20b"
context_window = 131072
logical_model = "openai-gpt-oss-20b"
equivalence_group = "openai-gpt-oss-20b"
served_variant = "nvidia-nim"
api_dialect = "openai_chat"
capabilities = ["tools", "streaming", "thinking"]
architecture = { parameter_count_b = 21.0, active_parameter_count_b = 3.6, moe = true, license = "Apache-2.0", source_url = "https://developers.openai.com/api/docs/models/gpt-oss-20b", last_verified = "2026-06-20" }
tier = "mid"
open_weight = true
strengths = ["speed", "cheap", "tool_use", "reasoning"]

# Removed 2026-09-20: nvidia/step-3.7-flash. `stepfun-ai/step-3.7-flash` is
# absent from https://integrate.api.nvidia.com/v1/models (observed 2026-09-20).
# The stepfun/step-3.7-flash route on OpenRouter is unaffected.

# Added 2026-09-20, observed on https://integrate.api.nvidia.com/v1/models the
# same day: NIM now serves Kimi K3 and both Z.ai GLM 5.3 builds. Per-token
# rates are not published on the NIM index, so these rows carry no `pricing`,
# like their Nemotron siblings above. Context windows and capabilities are
# inherited from the first-party rows (Moonshot's /v1/models reports 1,048,576
# for kimi-k3; the GLM rows follow the direct Z.AI rows): NIM publishes neither
# per-route limits nor a tool matrix, so treat both as unverified for NIM.
[models."nvidia/kimi-k3"]
name = "Kimi K3 (NVIDIA NIM)"
provider = "nvidia"
wire_model = "moonshotai/kimi-k3"
context_window = 1048576
logical_model = "moonshot-kimi-k3"
equivalence_group = "moonshot-kimi-k3"
served_variant = "nvidia-nim"
api_dialect = "openai_chat"
capabilities = ["tools", "vision", "streaming", "thinking"]
tier = "frontier"
open_weight = true
strengths = ["coding", "agentic", "long_context", "tool_use", "reasoning", "vision"]
complementary_with = ["anthropic-claude", "openai-gpt", "google-gemini", "qwen", "deepseek"]

[models."nvidia/glm-5.3"]
name = "GLM 5.3 (NVIDIA NIM)"
provider = "nvidia"
wire_model = "z-ai/glm-5.3"
context_window = 1048576
logical_model = "glm-5.3"
equivalence_group = "glm-5.3"
served_variant = "nvidia-nim"
api_dialect = "openai_chat"
capabilities = ["tools", "streaming", "thinking"]
tier = "frontier"
open_weight = true
strengths = ["coding", "agentic", "tool_use", "reasoning", "long_context"]

[models."nvidia/glm-5.3-flash"]
name = "GLM 5.3 Flash (NVIDIA NIM)"
provider = "nvidia"
wire_model = "z-ai/glm-5.3-flash"
context_window = 1048576
logical_model = "glm-5.3-flash"
equivalence_group = "glm-5.3-flash"
served_variant = "nvidia-nim"
api_dialect = "openai_chat"
capabilities = ["tools", "streaming", "thinking"]
# `frontier` matches the direct Z.AI and OpenRouter rows: tier is a property of
# the logical model, and every row in an equivalence group must agree.
tier = "frontier"
open_weight = true
strengths = ["speed", "cheap", "coding", "tool_use", "reasoning"]