v0.9.2

Provider v0.9.2 (Swift CLI)

Binary hash: 1451c571305e4411f576a40649cbaf6a9467acaa6381db61d1ffbbd359e65245
Bundle hash: 6bb5cf9f9433fc14ee78b3ba2c27778890198e55ac7e1a4cd9d53370196425d2
Metallib hash: 2129f6132794d84243c02631bc7478417dba0e0dbd10467beab56c11bf20bdd2 (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

Provider changes

• Gemma QAT SSD prefix caching — Enable authenticated complete paged checkpoints by default for exact gemma-4-26b-qat-4bit. Preserve tenant, model, prompt, binary, metallib and numerical-state identity checks, the global cache disable and cold fallback. Other Gemma artifacts and GPT-OSS caching remain opt-in. A new binary starts a new checkpoint identity; existing 0.9.1 checkpoints are not reused across the upgrade.
• Adaptive Gemma MTP — Automatically resolve the catalog assistant for that exact QAT target and select ordinary decode or one draft token from measured committed output and elapsed time. Include seed work, reset workload learning when requests finish or IDs are reused, and track compilation warmup by exact verification shape. Support target-prefix sampling for temperature/top-p/top-k/min-p, with ordinary decode for unsupported transforms and explicit diagnostic verification controls retained.
• Assistant activation while serving — Download and verify the optional assistant while the current engine serves. Reserve staging memory and retain its target against eviction; publish reduced capacity immediately and restore survivor KV grants after discarded preparation. After network rollout jitter, close only that model's new admissions, advertise reloading, and finish accepted work before swapping. Racing requests receive transient 503 slot_state refusals. Timeout or cancellation discards the candidate and reopens the original engine without cancelling accepted requests. Standalone follows the same bounded drain without fleet jitter; insufficient memory preserves target-only serving.
• Assistant download sources — Honor catalog-declared immutable Hugging Face assistant revisions with checksum-verified R2 fallback and jittered fetch retries. Existing R2-only metadata remains valid; shipping this binary does not apply the separate catalog patch. Standalone uses its configured coordinator catalog authority.
• Nemotron Lightning serving — Admit the three explicitly qualified registry/Hugging Face IDs on the existing network and standalone paths; reject other nemotron_h artifacts. Enable native paged KV and complete encrypted prefix reuse with native activation/KV precision and FP32 persistent Mamba state. Declared embedded MTP uses request-owned assistant state, exact prefix checkpoints and adaptive depth up to seven; checkpoints include Nemotron numerical controls.
• Native reasoning and tools — Separate Nemotron reasoning before tool parsing and validate required/named calls before publishing them through the existing encrypted response stream. Keep Gemma grammar enforcement and explicit per-model capability advertisement. Fix standalone Lightning admission to use the model's exact identity before the existing memory gate.
• Shared inference dependencies — Pin the merged MLX core, C, Swift and SDK chain for explicit mutable Metal-kernel inputs, request-owned paged MTP, recurrent rollback, checkpoint ownership and typed native generation events. Release CI includes nonzero/no-skip synthetic SDK gates; real-model gates remain separately identified.

This is a provider-only release. Coordinator warm-pool and console changes listed in the generated comparison require their own deployments. Validation and rollout notes.

What's Changed

• fix(console): retain earnings access after removing all Macs by @0xkydo in #888
• fix(registry): warm pool grows on measured headroom, not only after a failed request by @brandon-eigenlabs in #807
• Enable Nemotron Lightning in the coordinator-serving provider by @jonathan308 in #885
• feat(console-ui): payout-coverage notice on provider earnings page by @EigenMustafa in #887
• Fix standalone admission for qualified Nemotron Lightning models by @Gajesh2007 in #892
• Enable Gemma QAT prefix caching and adaptive MTP by @Gajesh2007 in #872
• chore(release): prepare provider v0.9.2 by @Gajesh2007 in #893

New Contributors

• @brandon-eigenlabs made their first contribution in #807
• @jonathan308 made their first contribution in #885
• @EigenMustafa made their first contribution in #887

Full Changelog: v0.9.1...v0.9.2