fix(cloudflare): correct catalogue reasoning efforts & base_models

providers/cloudflare-workers-ai/models/@cf/aisingapore/gemma-sea-lion-v4-27b-it.toml

@@ -1,23 +1,7 @@
+base_model = "aisingapore/gemma-sea-lion-v4-27b-it"
 name = "Gemma Sea Lion V4 27B It"
-description = "Open Gemma instruction model for efficient chat and self-hosted deployments"
-family = "gemma"
-release_date = "2025-09-23"
-last_updated = "2025-09-23"
-attachment = false
-reasoning = false
-temperature = true
-tool_call = false
 structured_output = false
-open_weights = true
 
 [cost]
 input = 0.351
 output = 0.555
-
-[limit]
-context = 128_000
-output = 128_000
-
-[modalities]
-input = ["text"]
-output = ["text"]

providers/cloudflare-workers-ai/models/@cf/deepseek-ai/deepseek-v4-flash-0731.toml

@@ -1,19 +1,20 @@
-# Toggle: thinking.type = enabled|disabled.
-# Effort: reasoning_effort = high|max.
+# Effort: reasoning_effort = none|low|high|max; none disables thinking.
+# Source: Workers AI /ai/models/search?format=openrouter (2026-09-18).
+# Context/output capped to Cloudflare Workers AI context window (1,048,576).
+# Search reports 1,310,720, exceeding AI Gateway max_completion_tokens.
+# This dated release adds a distinct low effort; the older preview does not.
+# Creator: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731#chat-template
 base_model = "deepseek/deepseek-v4-flash-0731"
 
-[[reasoning_options]]
-type = "toggle"
-
 [[reasoning_options]]
 type = "effort"
-values = ["high", "max"]
+values = ["none", "low", "high", "max"]
 
 [cost]
 input = 0.44
 output = 1.32
 cache_read = 0.014
 
 [limit]
-context = 1_310_720
+context = 1_048_576
 output = 1_048_576

providers/cloudflare-workers-ai/models/@cf/deepseek-ai/deepseek-v4-pro-0813.toml

@@ -1,8 +1,12 @@
-# Toggle: thinking.type = enabled|disabled.
-# Effort: reasoning_effort = high|max.
+# Effort: reasoning_effort = none|low|high|max; none disables thinking.
+# Source: Workers AI /ai/models/search?format=openrouter (2026-09-18).
+# This dated release adds a distinct low effort; the older preview does not.
+# Creator: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813#chat-template
 base_model = "deepseek/deepseek-v4-pro-0813"
 
-reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["high", "max"] }]
+[[reasoning_options]]
+type = "effort"
+values = ["none", "low", "high", "max"]
 
 [cost]
 input = 1.32

providers/cloudflare-workers-ai/models/@cf/google/gemma-4-26b-a4b-it.toml

@@ -1,11 +1,15 @@
-# Native `/ai/run` accepts `reasoning_effort = low|medium|high` and
-# `chat_template_kwargs.enable_thinking = true|false`; no budget is documented.
-# https://developers.cloudflare.com/workers-ai/models/gemma-4-26b-a4b-it/sync-input.json (accessed 2026-06-25)
+# Effort: reasoning_effort = none|high; none disables thinking.
+# Source: Workers AI /ai/models/search?format=openrouter (2026-09-18).
+# Per-model Search config normalizes low/medium to high; these are aliases,
+# not distinct levels despite the generic sync-input.json effort enum.
 
 base_model = "google/gemma-4-26b-a4b-it"
-reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high"] }]
 interleaved = true
 
+[[reasoning_options]]
+type = "effort"
+values = ["none", "high"]
+
 [cost]
 input = 0.1
 output = 0.3

providers/cloudflare-workers-ai/models/@cf/ibm-granite/granite-4.0-h-micro.toml

@@ -1,14 +1,7 @@
+base_model = "ibm/granite-4-h-micro"
 name = "Granite 4.0 H Micro"
 description = "Efficient model for low-latency assistance, extraction, and routine automation"
-family = "granite"
-release_date = "2025-10-07"
-last_updated = "2025-10-07"
-attachment = false
-reasoning = false
-temperature = true
-tool_call = true
 structured_output = false
-open_weights = true
 
 [cost]
 input = 0.017
@@ -17,7 +10,3 @@ output = 0.112
 [limit]
 context = 131_000
 output = 131_000
-
-[modalities]
-input = ["text"]
-output = ["text"]

providers/cloudflare-workers-ai/models/@cf/mistralai/mistral-small-3.1-24b-instruct.toml

@@ -1,23 +1,15 @@
+base_model = "mistral/mistral-small-3-1-24b-instruct-2503"
 name = "Mistral Small 3.1 24B Instruct"
 description = "Efficient Mistral model for fast chat, extraction, and production assistants"
-family = "mistral-small"
-release_date = "2025-03-18"
-last_updated = "2025-03-18"
 attachment = false
-reasoning = false
-temperature = true
-tool_call = true
 structured_output = false
-open_weights = true
 
 [cost]
 input = 0.351
 output = 0.555
 
 [limit]
-context = 128_000
 output = 128_000
 
 [modalities]
 input = ["text"]
-output = ["text"]

providers/cloudflare-workers-ai/models/@cf/moonshotai/kimi-k2.6.toml

@@ -1,9 +1,13 @@
-# Native `/ai/run` accepts `reasoning_effort = low|medium|high` and
-# `chat_template_kwargs.thinking = true|false`; no budget is documented.
-# https://developers.cloudflare.com/workers-ai/models/kimi-k2.6/sync-input.json (accessed 2026-06-25)
+# Effort: reasoning_effort = none|high; none selects instant mode.
+# Source: Workers AI /ai/models/search?format=openrouter (2026-09-18).
+# Per-model Search config normalizes low/medium to high; these are aliases,
+# not distinct levels despite the generic sync-input.json effort enum.
 
 base_model = "moonshotai/kimi-k2.6"
-reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high"] }]
+
+[[reasoning_options]]
+type = "effort"
+values = ["none", "high"]
 
 [interleaved]
 field = "reasoning_content"

providers/cloudflare-workers-ai/models/@cf/moonshotai/kimi-k2.7-code.toml

@@ -1,6 +1,8 @@
+# Thinking is mandatory; high is the only effective effort, so aliases add no caller control.
+# Source: Workers AI /ai/models/search?format=openrouter (2026-09-18).
 base_model = "moonshotai/kimi-k2.7-code"
-reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high"] }]
 temperature = true
+reasoning_options = []
 
 [cost]
 input = 0.95

providers/cloudflare-workers-ai/models/@cf/nvidia/nemotron-3-120b-a12b.toml

@@ -5,9 +5,15 @@
 base_model = "nvidia/nemotron-3-super-120b-a12b"
 name = "Nemotron 3 Super 120B"
 structured_output = true
-reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high"] }]
 interleaved = true
 
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "high"]
+
 [cost]
 input = 0.5
 output = 1.5

providers/cloudflare-workers-ai/models/@cf/qwen/qwen3.8-27b.toml

@@ -5,7 +5,13 @@
 
 base_model = "alibaba/qwen3.8-27b"
 description = "Qwen vision-language model for visual reasoning, documents, and agent tasks"
-reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "xhigh"] }]
+
+[[reasoning_options]]
+type = "toggle"
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "medium", "xhigh"]
 
 [cost]
 input = 0.45

providers/cloudflare-workers-ai/models/@cf/zai-org/glm-4.7-flash.toml

@@ -1,32 +1,21 @@
-# Native `/ai/run` accepts `reasoning_effort = low|medium|high` and
-# `chat_template_kwargs.enable_thinking = true|false`; no budget is documented.
-# https://developers.cloudflare.com/workers-ai/models/glm-4.7-flash/sync-input.json (accessed 2026-06-25)
-
-name = "GLM-4.7-Flash"
+# Toggle: chat_template_kwargs.enable_thinking = true|false (default: true).
+# Source: Workers AI /ai/models/search?format=openrouter (2026-09-18).
+# Workers AI's per-model config exposes only the binary thinking toggle;
+# the generic sync-input.json low/medium/high enum is not an effort selector here.
+# Creator template: https://huggingface.co/zai-org/GLM-4.7-Flash/blob/main/chat_template.jinja
+base_model = "zhipuai/glm-4.7-flash"
 description = "Efficient GLM model for fast reasoning, coding, and agent workflows"
-family = "glm-flash"
-release_date = "2026-01-19"
-last_updated = "2026-01-19"
-attachment = false
-reasoning = true
-reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high"] }]
-temperature = true
-tool_call = true
 structured_output = true
-knowledge = "2025-04"
-open_weights = true
 
 [interleaved]
 field = "reasoning_content"
 
+[[reasoning_options]]
+type = "toggle"
+
 [cost]
 input = 0.0605
 output = 0.4
 
 [limit]
 context = 131_072
-output = 131_072
-
-[modalities]
-input = ["text"]
-output = ["text"]

providers/cloudflare-workers-ai/models/@cf/zai-org/glm-5.2.toml

@@ -1,14 +1,13 @@
+# Effort: reasoning_effort = none|high|max; none disables thinking.
+# Source: Workers AI /ai/models/search?format=openrouter (2026-09-18).
 # Context and output limits verified against Cloudflare Workers AI's OpenAI-compatible endpoint.
 # https://developers.cloudflare.com/workers-ai/models/glm-5.2/
 base_model = "zhipuai/glm-5.2"
 name = "Glm 5.2"
 
-[[reasoning_options]]
-type = "toggle"
-
 [[reasoning_options]]
 type = "effort"
-values = ["low", "medium", "high"]
+values = ["none", "high", "max"]
 
 [cost]
 input = 1.4

providers/cloudflare-workers-ai/models/@cf/zai-org/glm-5.3-flash.toml

@@ -1,28 +1,23 @@
+# Effort: reasoning_effort = low|high|max; thinking is mandatory.
+# Source: Workers AI /ai/models/search?format=openrouter (2026-09-18).
 # Context/output capped to Cloudflare Workers AI context window (1,048,576).
-# Catalog/search previously reported 1,310,720, which exceeds AI Gateway max_completion_tokens.
-# https://developers.cloudflare.com/workers-ai/models/glm-5.3-flash/
+# Search reports 1,310,720, exceeding AI Gateway max_completion_tokens.
+base_model = "zhipuai/glm-5.3-flash"
 name = "Glm 5.3 Flash"
 description = "GLM vision model for visual reasoning, documents, and multimodal agents"
-family = "glm-flash"
-release_date = "2026-08-26"
-last_updated = "2026-08-26"
-attachment = true
-reasoning = true
-temperature = true
-tool_call = true
-structured_output = true
-open_weights = true
-reasoning_options = []
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
 
 [cost]
 input = 0.15
 output = 0.5
 cache_read = 0.03
 
 [limit]
-context = 1_310_720
+context = 1_048_576
 output = 1_048_576
 
 [modalities]
 input = ["text", "image"]
-output = ["text"]

providers/cloudflare-workers-ai/models/@cf/zai-org/glm-5.3.toml

@@ -1,13 +1,20 @@
+# Effort: reasoning_effort = low|high|max; thinking is mandatory.
+# Source: Workers AI /ai/models/search?format=openrouter (2026-09-18).
+# Context/output capped to Cloudflare Workers AI context window (1,048,576).
+# Search reports 1,310,720, exceeding AI Gateway max_completion_tokens.
 base_model = "zhipuai/glm-5.3"
 name = "Glm 5.3"
 description = "Flagship GLM model for hybrid reasoning, coding, and agentic engineering"
-reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high", "max"]
 
 [cost]
 input = 1.4
 output = 4.4
 cache_read = 0.26
 
 [limit]
-context = 1_310_720
-output = 1_310_720
+context = 1_048_576
+output = 1_048_576