diff --git a/providers/fireworks-ai/models/accounts/fireworks/models/glm-5p3-flash.toml b/providers/fireworks-ai/models/accounts/fireworks/models/glm-5p3-flash.toml index 7e8d208090a..8a1203370b8 100644 --- a/providers/fireworks-ai/models/accounts/fireworks/models/glm-5p3-flash.toml +++ b/providers/fireworks-ai/models/accounts/fireworks/models/glm-5p3-flash.toml @@ -1,14 +1,17 @@ # Pricing: https://fireworks.ai/models/fireworks/glm-5p3-flash (accessed 2026-09-02) # Released on Fireworks Serverless API 2026-08-26. -# Thinking-only model: reasoning cannot be disabled. Use the same two effective -# tiers as Fireworks GLM 5.3: low/medium collapse to high; max/xhigh select max. -# https://docs.fireworks.ai/api-reference/post-chatcompletions (accessed 2026-08-29) +# Thinking-only model: reasoning cannot be disabled. Three distinct effort +# tiers like Fireworks GLM 5.3 (NOT the two-tier GLM 5.2 collapse): verified +# empirically 2026-09-04 — same prompt yields ~9/17/559 reasoning tokens for +# low/high/max. max/xhigh select max. +# https://docs.fireworks.ai/api-reference/post-chatcompletions base_model = "zhipuai/glm-5.3-flash" name = "GLM 5.3 Flash" +last_updated = "2026-09-04" [[reasoning_options]] type = "effort" -values = ["high", "max"] +values = ["low", "high", "max"] [interleaved] field = "reasoning_content" diff --git a/providers/fireworks-ai/models/accounts/fireworks/models/glm-5p3.toml b/providers/fireworks-ai/models/accounts/fireworks/models/glm-5p3.toml index 8fb8037dd97..fa2c2de0072 100644 --- a/providers/fireworks-ai/models/accounts/fireworks/models/glm-5p3.toml +++ b/providers/fireworks-ai/models/accounts/fireworks/models/glm-5p3.toml @@ -1,16 +1,18 @@ # Pricing: https://docs.fireworks.ai/serverless/pricing (accessed 2026-08-29) # Released on Fireworks Serverless API 2026-08-28. # Thinking-only model: reasoning cannot be disabled (reasoning_effort="none" -# and thinking.type="disabled" both return 400). Two effective tiers like -# GLM 5.2: low/medium collapse to high, max/xhigh select max. -# https://docs.fireworks.ai/api-reference/post-chatcompletions (accessed 2026-08-29) +# and thinking.type="disabled" both return 400). Unlike GLM 5.2 (two tiers, +# low/medium collapse to high), GLM 5.3 has three distinct effort tiers: +# verified empirically 2026-09-04 — same prompt yields ~10/25/198 reasoning +# tokens for low/high/max. max/xhigh select max. +# https://docs.fireworks.ai/api-reference/post-chatcompletions base_model = "zhipuai/glm-5.3" name = "GLM 5.3" -last_updated = "2026-08-28" +last_updated = "2026-09-04" [[reasoning_options]] type = "effort" -values = ["high", "max"] +values = ["low", "high", "max"] [interleaved] field = "reasoning_content"