moltis-org/moltis

moltis-org

moltis

Blog Docs Changelog

Performance History

Latest Results

chore: prepare release 20261023.01

main

7 days ago

fix(providers): use canonical NEAR AI catalog model (#1318) NEAR AI now advertises z-ai/glm-5.3-flash as its canonical GLM entry and retains zai-org/GLM-5.1-FP8 only in metadata aliases. The old live catalog assertion failed even though chat requests through the alias still worked. Use the canonical ID in live checks and configuration examples, cover canonical discovery without alias duplication, and construct live providers through production registration. Keep the catalog membership assertion and validate text-generation capabilities rather than weakening the release gate.

main

7 days ago

fix(providers): use canonical NEAR AI catalog model NEAR AI now advertises z-ai/glm-5.3-flash as its canonical GLM entry and retains zai-org/GLM-5.1-FP8 only in metadata aliases. The old live catalog assertion failed even though chat requests through the alias still worked. Use the canonical ID in live checks and configuration examples, cover canonical discovery without alias duplication, and construct live providers through production registration. Keep the catalog membership assertion and validate text-generation capabilities rather than weakening the release gate.

fix/nearai-live-catalog

7 days ago

fix(providers): correct DeepSeek thinking tests and refresh Z.AI models (#1317) * fix(providers): refresh DeepSeek and Z.AI models and thinking requests DeepSeek live tests built bare OpenAI-compatible providers, bypassing the registered capability policy that enables thinking and replays reasoning. Use production registration in both live suites and verify the outgoing thinking fields and full assistant history with offline HTTP tests. Add the documented deepseek-flash and GLM-5.3/5.2 catalog entries and capabilities, and align reasoning effort with low/high/max. Keep the supported GLM-4.5-Flash regression alongside a current Flash smoke test. Preserve underlying stream transport causes without request URLs so connection failures are distinguishable from model or authentication errors. * fix(providers): match live reasoning and Z.AI capacity contracts The live API can issue a simple DeepSeek tool call with no reasoning tokens even when thinking is enabled. Check exact preservation of returned reasoning and successful tool-result replay instead of asserting a minimum amount of generated reasoning. Preserve explicitly empty reasoning without substituting visible text. Z.AI confirms quota and overload errors on the older free Flash pool while current GLM-5.3-Flash succeeds. Run the same behavioral contract against the current metered model, retain the free model in catalog health reporting, and preserve reasoning plus every tool result in multi-turn tests. * fix(providers): retain explicitly empty completion reasoning Keep an explicit empty reasoning_content distinct from an absent field so completion-based DeepSeek tool turns do not fall back to visible text. Verify the distinction through real registered HTTP providers; optional MiniMax replay still omits empty fields. * docs(config): keep Z.AI Coding Plan URL example valid

main

7 days ago

docs(config): keep Z.AI Coding Plan URL example valid

fix/deepseek-zai-live-gates

7 days ago

fix(providers): retain explicitly empty completion reasoning Keep an explicit empty reasoning_content distinct from an absent field so completion-based DeepSeek tool turns do not fall back to visible text. Verify the distinction through real registered HTTP providers; optional MiniMax replay still omits empty fields.

fix/deepseek-zai-live-gates

7 days ago

fix(providers): match live reasoning and Z.AI capacity contracts The live API can issue a simple DeepSeek tool call with no reasoning tokens even when thinking is enabled. Check exact preservation of returned reasoning and successful tool-result replay instead of asserting a minimum amount of generated reasoning. Preserve explicitly empty reasoning without substituting visible text. Z.AI confirms quota and overload errors on the older free Flash pool while current GLM-5.3-Flash succeeds. Run the same behavioral contract against the current metered model, retain the free model in catalog health reporting, and preserve reasoning plus every tool result in multi-turn tests.

fix/deepseek-zai-live-gates

7 days ago

fix(providers): refresh DeepSeek and Z.AI models and thinking requests DeepSeek live tests built bare OpenAI-compatible providers, bypassing the registered capability policy that enables thinking and replays reasoning. Use production registration in both live suites and verify the outgoing thinking fields and full assistant history with offline HTTP tests. Add the documented deepseek-flash and GLM-5.3/5.2 catalog entries and capabilities, and align reasoning effort with low/high/max. Keep the supported GLM-4.5-Flash regression alongside a current Flash smoke test. Preserve underlying stream transport causes without request URLs so connection failures are distinguishable from model or authentication errors.

fix/deepseek-zai-live-gates

7 days ago

Latest Branches

fix(providers): use canonical NEAR AI catalog model#1318

7 days ago

e1b6cf4

fix/nearai-live-catalog

-10%

fix(providers): correct DeepSeek thinking tests and refresh Z.AI models#1317

7 days ago

192e829

fix/deepseek-zai-live-gates

fix(providers): refresh retired MiniMax models and verify live tool use#1314

8 days ago

2a7d70d

fix/minimax-current-models

Home Terms Privacy Docs