feat: add Lyceum provider

models/alibaba/qwen3-30b-a3b-instruct-2507.toml

@@ -0,0 +1,21 @@
+# https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507 (accessed 2026-09-02)
+name = "Qwen3 30B-A3B Instruct 2507"
+description = "Qwen3 Mixture-of-Experts instruction model for multilingual chat, coding, and tool use"
+family = "qwen"
+release_date = "2025-07-29"
+last_updated = "2025-07-29"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+license = "apache-2.0"
+
+[limit]
+context = 262_144
+output = 16_384
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/alibaba/qwen3-embedding-8b.toml

@@ -0,0 +1,20 @@
+# https://huggingface.co/Qwen/Qwen3-Embedding-8B (accessed 2026-09-02)
+name = "Qwen3 Embedding 8B"
+description = "Text embedding model for semantic search, retrieval, clustering, and reranking"
+family = "text-embedding"
+release_date = "2025-06-05"
+last_updated = "2025-06-05"
+attachment = false
+reasoning = false
+temperature = false
+tool_call = false
+open_weights = true
+license = "apache-2.0"
+
+[limit]
+context = 32_768
+output = 0
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/google/gemma-3-27b-it.toml

@@ -0,0 +1,24 @@
+# Gemma 3 has no native function calling: the served endpoint accepts a tools
+# array but the model answers in prose instead of emitting tool_calls.
+# https://ai.google.dev/gemma/docs/core/model_card_3 (accessed 2026-09-02)
+name = "Gemma 3 27B IT"
+description = "Open multimodal Gemma instruction model for efficient chat, vision, and self-hosted deployments"
+family = "gemma"
+release_date = "2025-03-12"
+last_updated = "2025-03-12"
+attachment = true
+reasoning = false
+temperature = true
+tool_call = false
+structured_output = true
+knowledge = "2024-08"
+open_weights = true
+license = "gemma"
+
+[limit]
+context = 131_072
+output = 8_192
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

models/nousresearch/hermes-4-405b.toml

@@ -0,0 +1,21 @@
+# https://hermes4.nousresearch.com (accessed 2026-09-02)
+name = "Hermes 4 405B"
+description = "Hybrid reasoning model from Nous Research, built on Llama 3.1 405B for deliberate analysis, math, code, and tool use"
+family = "hermes"
+release_date = "2025-08-26"
+last_updated = "2025-08-26"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+license = "llama3.1"
+
+[limit]
+context = 131_072
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/nousresearch/hermes-4-70b.toml

@@ -0,0 +1,21 @@
+# https://hermes4.nousresearch.com (accessed 2026-09-02)
+name = "Hermes 4 70B"
+description = "Hybrid reasoning model from Nous Research, built on Llama 3.1 70B for deliberate analysis, math, code, and tool use"
+family = "hermes"
+release_date = "2025-08-26"
+last_updated = "2025-08-26"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+license = "llama3.1"
+
+[limit]
+context = 131_072
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/nvidia/cosmos3-super-reasoner.toml

@@ -0,0 +1,22 @@
+# Reasoning tower of NVIDIA Cosmos 3 Super, a 64B two-tower Mixture-of-Transformers
+# built on the Qwen3-VL 32B architecture.
+# https://developer.nvidia.com/blog/develop-physical-ai-reasoning-world-and-action-models-with-nvidia-cosmos-3/
+# (accessed 2026-09-02)
+name = "Cosmos3 Super Reasoner"
+description = "Physical-AI reasoning model for spatial understanding, multi-agent planning, and vision grounding"
+release_date = "2026-06"
+last_updated = "2026-06"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+
+[limit]
+context = 256_000
+output = 256_000
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

models/openbmb/minicpm-v-4_5.toml

@@ -0,0 +1,20 @@
+# 8.7B vision-language model with controllable hybrid fast/deep thinking.
+# https://huggingface.co/openbmb/MiniCPM-V-4_5 (accessed 2026-09-02)
+name = "MiniCPM-V 4.5"
+description = "Compact vision-language model for OCR, document parsing, and video understanding"
+release_date = "2025-08"
+last_updated = "2025-09"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = false
+open_weights = true
+license = "apache-2.0"
+
+[limit]
+context = 32_768
+output = 32_768
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]

providers/lyceum/models/deepseek/deepseek-v4-flash-0731.toml

@@ -0,0 +1,14 @@
+# Effort: reasoning_effort = none|low|high|max; none disables thinking.
+# The unlisted values are accepted but collapse onto these, as on DeepSeek's own API.
+base_model = "deepseek/deepseek-v4-flash-0731"
+reasoning_options = [{ type = "effort", values = ["none", "low", "high", "max"] }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.25
+output = 0.3
+
+[limit]
+output = 65_536

providers/lyceum/models/deepseek/deepseek-v4-pro.toml

@@ -0,0 +1,12 @@
+# Effort: reasoning_effort = none|high|max; none disables thinking.
+# low and medium map to high upstream, and minimal skips thinking like none.
+base_model = "deepseek/deepseek-v4-pro"
+reasoning_options = [{ type = "effort", values = ["none", "high", "max"] }]
+interleaved = true
+
+[cost]
+input = 1.75
+output = 3.5
+
+[limit]
+output = 65_536

providers/lyceum/models/google/gemma-3-27b-it.toml

@@ -0,0 +1,5 @@
+base_model = "google/gemma-3-27b-it"
+
+[cost]
+input = 0.1
+output = 0.3

providers/lyceum/models/lyceum/complex.toml

@@ -0,0 +1,18 @@
+# Router alias resolved server-side to minimax/minimax-m3; controls and price are its own.
+# No caller control: reasoning_effort, enable_thinking, thinking, thinking.type,
+# reasoning.enabled and chat_template_kwargs.enable_thinking all leave the
+# reasoning on.
+base_model = "minimax/MiniMax-M3"
+name = "Lyceum Complex"
+description = "Routing alias that Lyceum resolves to minimax/minimax-m3"
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.4
+output = 2
+
+[limit]
+output = 65_536

providers/lyceum/models/lyceum/reasoning.toml

@@ -0,0 +1,15 @@
+# Router alias resolved server-side to z-ai/glm-5.2; controls and price are its own.
+# Effort: reasoning_effort = none|high|max; none disables thinking.
+# low and medium map to high upstream, and minimal skips thinking like none.
+base_model = "zhipuai/glm-5.2"
+name = "Lyceum Reasoning"
+description = "Routing alias that Lyceum resolves to z-ai/glm-5.2"
+reasoning_options = [{ type = "effort", values = ["none", "high", "max"] }]
+interleaved = true
+
+[cost]
+input = 1.5
+output = 4.5
+
+[limit]
+output = 65_536

providers/lyceum/models/lyceum/simple.toml

@@ -0,0 +1,13 @@
+# Router alias resolved server-side to qwen/qwen3.5-9b; controls and price are its own.
+# Toggle: reasoning_effort = none disables thinking, any other accepted value enables it.
+base_model = "alibaba/qwen3.5-9b"
+name = "Lyceum Simple"
+description = "Routing alias that Lyceum resolves to qwen/qwen3.5-9b"
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.15
+output = 0.2

providers/lyceum/models/meta-llama/llama-3.3-70b-instruct.toml

@@ -0,0 +1,8 @@
+# The lab row carries attachment = true next to text-only modalities; this host
+# rejects an image_url content part outright, so the override is a real delta.
+base_model = "meta/llama-3.3-70b-instruct"
+attachment = false
+
+[cost]
+input = 0.13
+output = 0.4

providers/lyceum/models/minimax/minimax-m2.5.toml

@@ -0,0 +1,16 @@
+# Always reasons. reasoning_effort = none is rejected, and no other wire control
+# turns it off: enable_thinking, thinking, thinking.type, reasoning.enabled and
+# chat_template_kwargs.enable_thinking all leave it on. reasoning.exclude = true
+# only hides the reasoning, which still burns the whole max_tokens budget.
+base_model = "minimax/MiniMax-M2.5"
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.3
+output = 1.2
+
+[limit]
+output = 65_536

providers/lyceum/models/minimax/minimax-m3.toml

@@ -0,0 +1,15 @@
+# No caller control: reasoning_effort, enable_thinking, thinking, thinking.type,
+# reasoning.enabled and chat_template_kwargs.enable_thinking all leave the
+# reasoning on.
+base_model = "minimax/MiniMax-M3"
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.4
+output = 2
+
+[limit]
+output = 65_536

providers/lyceum/models/moonshotai/kimi-k2.6.toml

@@ -0,0 +1,13 @@
+# Toggle: reasoning_effort = none disables thinking, any other accepted value enables it.
+base_model = "moonshotai/kimi-k2.6"
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 1
+output = 4
+
+[limit]
+output = 65_536

providers/lyceum/models/moonshotai/kimi-k2.7-code.toml

@@ -0,0 +1,15 @@
+# No caller control: reasoning_effort, enable_thinking, thinking, thinking.type,
+# reasoning.enabled and chat_template_kwargs.enable_thinking all leave the
+# reasoning on.
+base_model = "moonshotai/kimi-k2.7-code"
+reasoning_options = []
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 1.25
+output = 4.5
+
+[limit]
+output = 65_536

providers/lyceum/models/moonshotai/kimi-k3.toml

@@ -0,0 +1,12 @@
+# Effort: reasoning_effort = none|low|high|max; none disables thinking.
+# medium is rejected upstream; the rest collapse onto these levels.
+base_model = "moonshotai/kimi-k3"
+reasoning_options = [{ type = "effort", values = ["none", "low", "high", "max"] }]
+interleaved = true
+
+[cost]
+input = 3
+output = 15
+
+[limit]
+output = 65_536

providers/lyceum/models/nousresearch/hermes-4-405b.toml

@@ -0,0 +1,11 @@
+# Hybrid reasoner driven by the system prompt, not a request field: it emits
+# <think> blocks inside content, and reasoning_effort changes nothing.
+base_model = "nousresearch/hermes-4-405b"
+reasoning_options = []
+
+[cost]
+input = 1
+output = 3
+
+[limit]
+output = 65_536

providers/lyceum/models/nousresearch/hermes-4-70b.toml

@@ -0,0 +1,11 @@
+# Hybrid reasoner driven by the system prompt, not a request field: it emits
+# <think> blocks inside content, and reasoning_effort changes nothing.
+base_model = "nousresearch/hermes-4-70b"
+reasoning_options = []
+
+[cost]
+input = 0.13
+output = 0.4
+
+[limit]
+output = 65_536

providers/lyceum/models/nvidia/cosmos3-super-reasoner.toml

@@ -0,0 +1,11 @@
+# Reasons into <think> blocks inside content when the system prompt asks for it;
+# reasoning_effort is accepted but changes nothing.
+base_model = "nvidia/cosmos3-super-reasoner"
+reasoning_options = []
+
+[cost]
+input = 0.1
+output = 0.3
+
+[limit]
+output = 65_536

providers/lyceum/models/nvidia/llama-3_1-nemotron-ultra-253b-v1.toml

@@ -0,0 +1,9 @@
+# Reasoning is switched by the "detailed thinking on" system prompt, not by a
+# request field: the <think> block arrives inside content and reasoning_effort
+# changes nothing.
+base_model = "nvidia/llama-3.1-nemotron-ultra-253b"
+reasoning_options = []
+
+[cost]
+input = 0.6
+output = 1.8