88# - structured_output: a quality judgement. Cloudflare advertises response_format for
99# nearly every model, but several do not honour it (verified live: qwq-32b, qwen2.5-
1010# coder, gemma-sea-lion return non-conforming output), so this stays hand-set.
11- # - reasoning_options: proxied only (the catalog has no reasoning schema).
12- # - limit: proxied context/output that deliberately differ from the catalog.
11+ # - reasoning_options: proxied (catalog has no reasoning schema) and any hosted @cf model
12+ # whose docs schema doesn't describe its real reasoning knob (native, non-graded, etc.).
13+ # - interleaved: whether a wire returns reasoning content as a separate field. Never
14+ # inheritable from base_model (this deployment's wire format can differ from the lab's).
15+ # - limit.output: neither the catalog's max_output_tokens nor the docs schema is a reliable
16+ # ceiling (both verified wrong against lab/first-party for several models), so output is
17+ # always hand-curated; only limit.context is auto-derived.
18+ # - note: leading `#` comment lines (wire-path docs, source citations) that would otherwise
19+ # be lost on every regeneration.
1320# Consumed by: bun run cloudflare-ai-gateway:generate
1421
1522# Proxied Text-Generation catalog ids with no lab file under models/ (cannot publish yet).
@@ -115,8 +122,26 @@ reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
115122# Proxied reasoning models whose reasoning shape the Cloudflare catalog schema does not
116123# expose (native-format providers). Values mirror models.dev's merge-gateway provider, which
117124# proxies the same upstreams. Google uses thinkingConfig.thinkingBudget / thinking_level.
125+ #
126+ # DeepSeek: toggle-only would only be correct if AI Gateway drops the effort field on this
127+ # route (merge-gateway's own toggle-only shape is backed by a live-tested selected-route
128+ # comment specific to *their* relay, not evidence about Cloudflare's). Absent host-specific
129+ # proof that Cloudflare strips it, match the lab + Workers AI baseline: toggle + high/max.
118130[models ."deepseek/deepseek-v4-pro" ]
119- reasoning_options = [{ type = " toggle" }]
131+ note = [" Toggle: thinking.type = enabled|disabled." , " Effort: reasoning_effort = high|max." ]
132+ reasoning_options = [{ type = " toggle" }, { type = " effort" , values = [" high" , " max" ] }]
133+
134+ # Alibaba is a first-party lab; the catalog's OpenAI-compat schema exposes a generic
135+ # `reasoning_effort` enum, but Alibaba's real API (see providers/alibaba/models/qwen3.7-plus.toml)
136+ # is toggle + budget_tokens, not graded effort. Match the lab shape per AGENTS.md.
137+ [models ."alibaba/qwen3 .7-plus" ]
138+ note = [" Toggle: enable_thinking true|false." , " Budget: thinking_budget (integer reasoning tokens)." ]
139+ reasoning_options = [{ type = " toggle" }, { type = " budget_tokens" }]
140+
141+ # First-party xAI Grok 4.6 (providers/xai/models/grok-4.6.toml) includes xhigh; the catalog
142+ # schema enum is missing it.
143+ [models ."xai/grok-4 .6" ]
144+ reasoning_options = [{ type = " effort" , values = [" low" , " medium" , " high" , " xhigh" ] }]
120145
121146[models ."google/gemini-2 .5-pro" ]
122147reasoning_options = [{ type = " budget_tokens" , min = 128 , max = 32_768 }]
@@ -157,39 +182,55 @@ base_model = "deepseek/deepseek-r1-distill-qwen-32b"
157182structured_output = false
158183# Always-on reasoner with no user-facing knob; docs schema exposes no reasoning parameter.
159184reasoning_options = []
185+ limit = { output = 80_000 }
160186
161187[models ."workers-ai/@cf/deepseek-ai/deepseek-v4-flash-0731" ]
162188base_model = " deepseek/deepseek-v4-flash-0731"
163189
164190[models ."workers-ai/@cf/deepseek-ai/deepseek-v4-pro-0813" ]
165191base_model = " deepseek/deepseek-v4-pro-0813"
192+ note = [" Toggle: thinking.type = enabled|disabled." , " Effort: reasoning_effort = high|max." ]
193+ reasoning_options = [{ type = " toggle" }, { type = " effort" , values = [" high" , " max" ] }]
166194
167195[models ."workers-ai/@cf/google/gemma-4-26b-a4b-it" ]
168196base_model = " google/gemma-4-26b-a4b-it"
197+ note = [
198+ " Native `/ai/run` accepts `reasoning_effort = low|medium|high` and" ,
199+ " `chat_template_kwargs.enable_thinking = true|false`; no budget is documented." ,
200+ " https://developers.cloudflare.com/workers-ai/models/gemma-4-26b-a4b-it/sync-input.json (accessed 2026-06-25)" ,
201+ ]
202+ interleaved = true
203+ limit = { output = 16_384 }
169204
170205[models ."workers-ai/@cf/ibm-granite/granite-4 .0-h-micro" ]
171206base_model = " ibm/granite-4-h-micro"
172207structured_output = false
208+ limit = { output = 131_000 }
173209
174210[models ."workers-ai/@cf/meta/llama-3 .1-8b-instruct-fp8" ]
175211base_model = " meta/llama-3.1-8b-instruct"
176212structured_output = false
213+ limit = { output = 32_000 }
177214
178215[models ."workers-ai/@cf/meta/llama-3 .2-11b-vision-instruct" ]
179216base_model = " meta/llama-3.2-11b-vision-instruct"
180217structured_output = false
218+ limit = { output = 128_000 }
181219
182220[models ."workers-ai/@cf/meta/llama-3 .2-1b-instruct" ]
183221base_model = " meta/llama-3.2-1b"
184222structured_output = false
223+ limit = { output = 60_000 }
185224
186225[models ."workers-ai/@cf/meta/llama-3 .2-3b-instruct" ]
187226base_model = " meta/llama-3.2-3b"
188227structured_output = false
228+ limit = { output = 80_000 }
189229
190230[models ."workers-ai/@cf/meta/llama-3 .3-70b-instruct-fp8-fast" ]
191231base_model = " meta/llama-3.3-70b-instruct"
192232structured_output = false
233+ limit = { output = 24_000 }
193234
194235[models ."workers-ai/@cf/meta/llama-4-scout-17b-16e-instruct" ]
195236base_model = " meta/llama-4-scout-17b-instruct"
@@ -198,40 +239,73 @@ structured_output = false
198239[models ."workers-ai/@cf/meta/llama-guard-3-8b" ]
199240base_model = " meta/llama-guard-3-8b"
200241structured_output = false
242+ limit = { output = 131_072 }
201243
202244[models ."workers-ai/@cf/mistralai/mistral-small-3 .1-24b-instruct" ]
203245base_model = " mistral/mistral-small-3-1-24b-instruct-2503"
204246structured_output = false
247+ limit = { output = 128_000 }
205248
206249[models ."workers-ai/@cf/moonshotai/kimi-k2 .6" ]
207250base_model = " moonshotai/kimi-k2.6"
251+ note = [
252+ " Native `/ai/run` accepts `reasoning_effort = low|medium|high` and" ,
253+ " `chat_template_kwargs.thinking = true|false`; no budget is documented." ,
254+ " https://developers.cloudflare.com/workers-ai/models/kimi-k2.6/sync-input.json (accessed 2026-06-25)" ,
255+ ]
256+ interleaved = { field = " reasoning_content" }
257+ limit = { output = 256_000 }
208258
209259[models ."workers-ai/@cf/moonshotai/kimi-k2 .7-code" ]
210260base_model = " moonshotai/kimi-k2.7-code"
211261
212262[models ."workers-ai/@cf/nvidia/nemotron-3-120b-a12b" ]
213263base_model = " nvidia/nemotron-3-super-120b-a12b"
214264structured_output = true
265+ note = [
266+ " Native `/ai/run` accepts `reasoning_effort = low|medium|high` and" ,
267+ " `chat_template_kwargs.enable_thinking = true|false`; no budget is documented." ,
268+ " https://developers.cloudflare.com/workers-ai/models/nemotron-3-120b-a12b/sync-input.json (accessed 2026-06-25)" ,
269+ ]
270+ interleaved = true
271+ limit = { output = 256_000 }
215272
216273[models ."workers-ai/@cf/openai/gpt-oss-120b" ]
217274base_model = " openai/gpt-oss-120b"
275+ note = [
276+ " Native `/ai/run` schema has no reasoning control, and no model-specific" ,
277+ " OpenAI-compatible reasoning field was verified. No token budget is documented." ,
278+ " https://developers.cloudflare.com/workers-ai/models/gpt-oss-120b/sync-input.json (accessed 2026-06-25)" ,
279+ ]
280+ reasoning_options = []
281+ limit = { output = 16_384 }
218282
219283[models ."workers-ai/@cf/openai/gpt-oss-20b" ]
220284base_model = " openai/gpt-oss-20b"
285+ note = [
286+ " Native `/ai/run` schema has no reasoning control, and no model-specific" ,
287+ " OpenAI-compatible reasoning field was verified. No token budget is documented." ,
288+ " https://developers.cloudflare.com/workers-ai/models/gpt-oss-20b/sync-input.json (accessed 2026-06-25)" ,
289+ ]
290+ reasoning_options = []
291+ limit = { output = 16_384 }
221292
222293[models ."workers-ai/@cf/qwen/qwen2 .5-coder-32b-instruct" ]
223294base_model = " alibaba/qwen2.5-coder-32b-instruct"
224295structured_output = false
296+ limit = { output = 32_768 }
225297
226298[models ."workers-ai/@cf/qwen/qwen3-30b-a3b-fp8" ]
227299base_model = " alibaba/qwen3-30b-a3b"
228300structured_output = false
229301reasoning_options = []
302+ limit = { output = 32_768 }
230303
231304[models ."workers-ai/@cf/qwen/qwq-32b" ]
232305base_model = " alibaba/qwq-32b"
233306structured_output = false
234307reasoning_options = []
308+ limit = { output = 24_000 }
235309
236310# structured_output inherited (base=true, verified live: honours response_format json_schema);
237311# reasoning_options (toggle + effort) derived from the docs input schema.
@@ -241,7 +315,11 @@ base_model = "alibaba/qwen3.8-27b"
241315[models ."workers-ai/@cf/zai-org/glm-4 .7-flash" ]
242316base_model = " zhipuai/glm-4.7-flash"
243317structured_output = true
318+ interleaved = { field = " reasoning_content" }
319+ limit = { output = 131_072 }
244320
245321[models ."workers-ai/@cf/zai-org/glm-5 .2" ]
246322base_model = " zhipuai/glm-5.2"
323+ note = [" Context and output limits verified against Cloudflare Workers AI's OpenAI-compatible endpoint." ]
324+ limit = { output = 256_000 }
247325
0 commit comments