ZeroLog
Private preview
EN DE

Product

10 November: more Featherless models answer instead of refusing

Several Featherless models that turned requests down or went silent now answer, and DeepSeek tool calls no longer leave markup in the reply.

User impact: Targeted change

People who chat with less common Featherless models get answers where several models refused or went silent, and DeepSeek V4.1 Flash no longer leaves tags in its replies. No action is needed.

Refusals that arrive late. Featherless sometimes reports a refused request inside the stream, after it has already signalled success. ZeroLog took such a reply as finished and showed the error. These refusals are now recognised, and the chat retries as it already did for a direct refusal: with a stricter message format, with tools described as text for models that reject native tools, without the reasoning effort for models that refuse that field, or with a smaller limit. Affected models included Devstral, Magistral, GLM-4.6, Kimi K2 Instruct, Apertus and several Qwen3 models.

Models that vanish while loading. A few less common models first report that they are loading and then that they are not deployed. The chat now moves on to the fallback model with a notice instead of ending with an error.

Tool calls. Kimi K2.6 and MiMo V2.5 have trouble with tool names that contain a dot. Their calls now reach the right tool. DeepSeek V4.1 Flash sometimes mixes two call formats: the calls ran, but leftover tags stayed in the answer and the next round came back empty. The tags are now removed, and an empty round is retried once with the call format spelled out.

No action is needed. If a Featherless chat ended with "The request was rejected as invalid" on 9 or 10 November, send the message again.

Back to Newsroom