Release Notes

Release 2.0.9-variant

This release contains model updates outlined in the following sections.

Kimi-K2.6

This is an updated release of Kimi-K2.6. This NIM is part of the NIM Certified offering. For more information on this model, refer to the model card.

For GPU support, refer to the support matrix for Kimi-K2.6.

Note the following limitations:

• Kimi-K2.6

  • Air-gapped and offline deployment is not supported in this release. The NIM contacts NGC at startup to download the speculative-decoding draft model, even when the model cache is fully pre-populated, so a deployment without network access to NGC will not start.

  • Structured (guided) decoding is partially supported. Only the regex mode is reliable. The choice, json, grammar, and whitespace_pattern modes may produce malformed output, such as extra or duplicated characters at the start of the response.

Muse Glimmer

This is the initial release of Muse Glimmer. For more information on this model, refer to the model card.

For GPU support, refer to the support matrix for Muse Glimmer.

Note the following limitations:

• Muse Glimmer

  • This model does not support disabling thinking. Thinking is always enabled when using the /v1/messages endpoint.

  • Formatting can be unstable, which may cause errors when extracting the final answer.

  • Chain-of-thought content is returned in the message.reasoning field. The final answer is returned in the message.content field.

  • If max_tokens is set too low, the response can return content: null with finish_reason: "length". This indicates an insufficient token budget, not an error. For the max_tokens field, use at least 512 for typical chat requests and at least 4096 for tool-calling requests.

  • Parallel tool calling can sometimes fall back to a single sequential tool call (observed in about 20% of cases).

  • Do not configure <|eom|> (token ID 200007) as a stop token under any circumstances.

  • DFlash can be enabled on all GPUs. To enable it, select a profile whose name ends with -dflash. min_p and logit_bias are not supported on DFlash profiles.

  • Video input requires mounting FFmpeg8 into the container. Add the following volume mount to your docker run command:

    -v /path/to/ffmpeg8/lib:/opt/ffmpeg8:ro \
    

  • The container docker run command does not support the -u $(id -u) parameter.

  • NIM_DISABLE_MODEL_DOWNLOAD is not supported.

  • For air-gap deployment, mount the local model path and set NIM_MODEL_PATH:

    -v <local-model-path>:/opt/nim/workspace \
    -e NIM_MODEL_PATH=/opt/nim/workspace
    

For information about past updates and older versions, refer to the previous release notes.