v0.9.6

Highlights

• Add native Bonsai 2 27B runtime support for text, vision, tools and paged KV, preserving the published checkpoint and its native precision.
• Reduce retained recurrent prefill buffers and apply the serving memory guard to throughput benchmarks.
• Fix the SSD write-behind shutdown/drain race that blocked the earlier 0.9.6 candidate.
• Validate sealed-app Qwen Metal resource lookup during packaging and runtime smoke.

Model availability: this release prepares provider support for Bonsai 2; its production catalog activation is a separate rollout. Qwen 3.8 Flash and Qwen3-VL remain retired from the production catalog.

Provider v0.9.6 (Swift CLI)

Binary hash: 5fa82269804aa832a0cb3fcb367f037d22b47652e8b3670f2a239231b9ab4462
CodeDirectory SHA-256: 75f5deb7745c616dad51512632e825f06ed4ccd22974797ed7ca89a90d8d159f
Bundle hash: 96575798ef1a4738d077657fa34366097272a0fa395e7a660775f4898bb9a79a
Metallib hash: 38ceb8a1113b373ccaa89355d895fc8ded3bafc81076dc4ffdb426e2eee03a93 (built from mlx 3fa8f25e6451)
Signed by: Developer ID Application: Eigen Labs, Inc. (SLDQ2GJ6TL)
Notarized: yes
Min macOS: 14.0

Install

curl -fsSL https://api.darkbloom.dev/install.sh | bash

What's Changed

• feat(console): add NVIDIA branding for Nemotron by @prasiddhnaik in #1117
• Fix Flash-Next signed-app resource crash for provider 0.9.6 by @Gajesh2007 in #1118
• fix: wait for SSD pipeline teardown before completing drain by @Gajesh2007 in #1119
• Bonsai 2 27B: native packed model, vision and tool integration by @jonathan308 in #1124

Full Changelog: v0.9.5...v0.9.6