Muna Server 0.0.23

• Added compute_architecture to heartbeat node status, so the control plane can discover which models have implementations for the node's GPUs.
• Added devices to heartbeat goal descriptors: the node loads the model's engine on the given GPUs, and reloads it when the assignment changes.
• Added policy to heartbeat goal descriptors, the per-model inputs to the local lease-weight solver.
• Added device_ids and build to heartbeat model status.
• Updated device-lease weights to be solved locally every second from live lease demand and each model's policy. Models without a policy are weighted equally.
• Updated the load-time VRAM measurement to cover only the model's assigned GPUs.
• Upgraded muna to 0.0.34.

Muna Server 0.0.22

• Added support for Anthropic image blocks in /v1/messages, with base64 or url sources.
• Updated the request body limit from 2 MB to 64 MB to accommodate base64-encoded images.
• Updated oversized request bodies to return 413 with a request_too_large error in the OpenAI or Anthropic envelope, instead of 400.
• Updated /v1/chat/completions and /v1/messages to fetch remote images on the server and pass them to the model inline.
• Upgraded muna to 0.0.29.

Muna Server 0.0.21

• Fixed /v1/messages (Anthropic client) returning 400 when a request replays thinking blocks from a prior response.
• Updated /v1/messages to return a descriptive 400 for unsupported image, document, and unrecognized content blocks.
• Upgraded muna to 0.0.28.

Muna Server 0.0.20

• Added max_input_tokens, max_tokens, and capabilities to /v1/models rows, derived from the predictor signature, satisfying the Anthropic API shape.
• Added GET /v1/models/{tag} for retrieving a single model.
• Upgraded muna to 0.0.27.

Muna Server 0.0.19

• Added support for thinking and output_config.effort in /v1/messages (Anthropic client).
• Added support for temperature, top_p, seed, and reasoning_effort in /v1/chat/completions (OpenAI client).
• Upgraded muna to 0.0.26.

Muna Server 0.0.18

• Upgraded muna to 0.0.25.

Muna Server 0.0.17

• Process-tier model downloads now run ahead of disk-tier prefetch, so a model the control plane wants loaded is ready sooner on nodes with slow disks.
• Resource downloads are published as soon as they are in memory instead of waiting for the disk to absorb them allowing for faster model cold starts.
• Upgraded muna to 0.0.24.

Muna Server 0.0.16

• Upgraded muna to 0.0.23.

Muna Server 0.0.15

• Updated streaming chat completions and messages endpoints to propagate input errors with 400 status code.
• Upgraded muna to 0.0.22.

Muna Server 0.0.14

• Fixed a sporadic segmentation fault when two requests to a static or dynamically batched model overlapped.