{"version":"0.10.6","platform":"macos-arm64","backend":"mlx-swift","binary_hash":"4474b0605343b502f58c45e7e696dba575d4802eb56cd30802ab9b526371c7c3","bundle_hash":"d519a55c2d6fa82fdc84f99ad63afb987c0210fc730b1ba204c30c8dcd9ac714","metallib_hash":"98bab9d77dc5078ad8dd8f1fd17e8f3581415193506e4791a3c4f1ebaab614be","url":"https://pub-3d1cb668259340eeb2276e1d375c846d.r2.dev/releases/v0.10.6/artifacts/d519a55c2d6fa82fdc84f99ad63afb987c0210fc730b1ba204c30c8dcd9ac714/darkbloom-bundle-macos-arm64.tar.gz","changelog":"v0.10.6 - Two-Mac clusters, GLM-5.3-Flash and Qwen3.8-27B in tensor parallel, cluster runtime facts in each release\n\nServe one model on two Apple silicon Macs joined by one Thunderbolt 5 cable: run darkbloom cluster on each Mac (macOS 26.2 or later, RDMA over Thunderbolt turned on, the same Darkbloom release on both) and pick a model. The planner chooses tensor parallel or a pipeline; GLM-5.3-Flash and Qwen3.8-27B can serve in tensor parallel, with images and video. The cluster copies a model the other Mac lacks over the cable and comes back after a reboot. When one Mac goes away, the other keeps its part ready for 10 minutes and then serves alone only if it can hold the model, and the pair re-forms when the Mac returns. Prompt prefixes are kept per tenant in memory and in the encrypted SSD tier. darkbloom cluster local serves a pair without a coordinator.\n\nEvery release lane computes the build's cluster runtime facts from the final signed app, and this release registers them in production: for each model it qualifies on a pair and the production catalog lists, the digests a pair checks and its pair-serving grant, if any, and for GLM-5.3-Flash and Qwen3.8-27B public qualifications in pipeline and tensor parallel. A coordinator with darkbloom-platform#643 and its release-approval switch on derives the build's cluster approvals from them; every other model's pairs serve their owner only. A pair without a cluster approval serves only on its leading Mac's local API.\n\nA registered pair's cluster service is no longer restarted by its own part's heartbeats, pair members register the models their pair-serving grant admits, and a single Qwen3.8-27B request in tensor parallel decodes about 6% faster on two M5 Ultra. darkbloom models download and the darkbloom start picker show one progress line.\n\nOn one Mac, a content part of an unknown type is refused with 400, media the model's processor refuses answers 400 invalid_media instead of the retryable 503, and a response_format answer cut off by max_tokens is a client error that the coordinator no longer retries on another provider. The local API also normalizes requests as the coordinator path does, returns logprobs and reasoning-token counts, and says when max_tokens cut a response_format answer off. provider.toml is written owner-only, and the updater extracts releases under umask 022. Single-Mac serving needs no coordinator change or provider-version floor. libs/mlx (72dffa2d) and libs/mlx-swift (c633d221) are unchanged; libs/mlx-swift-lm moves to 7933a0fa for SDK #373, an M5 Ultra layout for the cluster's Qwen tensor-parallel projections.\n\nAlso in 0.10.6:\n- A co-resident load funds every model's KV floor before the proportional split, so models such as gemma-4-26b-qat-4bit and gpt-oss-20b are no longer refused beside Nemotron 3.5 Lightning when the floors fit.\n- A Qwen3.8-27B pair set up with darkbloom cluster plans without the local-only tensor assistant, so it registers with the coordinator and can serve public traffic.\n- A coordinator-side member end no longer counts toward taking a pair off public traffic, and a failing set-aside lasts one hour.\n- The release facts step leaves out another model's catalog ID for a shared artifact.\n- Pair Macs update with darkbloom update on both Macs, then a logout and login.","active":true,"created_at":"2026-10-31T16:41:50.633815Z"}