v0.10.0

Provider v0.10.0 (Swift CLI)

Source: 1ef9c89
Build and SDK qualification: https://github.com/Layr-Labs/d-inference/actions/runs/39876150189
Signing: https://github.com/Layr-Labs/d-inference/actions/runs/39876150189
Binary SHA-256: ae2e1ba1957236e20cf8e5d38ba9018c34decdb5705bbc13e441e1c8495980da
CodeDirectory SHA-256: d34117581ad5f9396fb3e68280ef33d5d5c5c79a91faa09c68f1ce7ffeaa0427
Bundle SHA-256: e04a6a6fb3cc3b667739570b1a04d2bfe2022fc554b52136c7668dab4b83629e
Metallib SHA-256: 156f2624f9f1b1413dfabd8b6e16423121b0e5b18e2573c23c4e023b606a462d
Signed and notarized: yes
macOS deployment target: 14.0 (build, tests and signed runtime smoke: macOS 27)

Install

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

What's Changed

• Separate model registration from R2 uploads and document the Bonsai incident by @Gajesh2007 in #1514
• devnet: dev-branch build lane, staging environment, and pre-release installer path by @davidtai in #1513
• release: publish dev and staging on master push (workflow_call refactor) by @davidtai in #1515
• fix(ci): preserve threat-review context and diagnose failures by @antojoseph in #1522
• Default deeper complete-checkpoint SSD restores on large-memory Macs by @javiech1 in #1520
• fix(provider): reject Qwen histories without a user query as a client 400 by @javiech1 in #1524
• fix(provider): reconcile interrupted watchdog launch receipts by @darkbloombugs in #1394
• docs(provider): qualify model-fit diagnostic contract by @darkbloombugs in #1434
• fix(provider): preserve refused watchdog kickstart receipts by @darkbloombugs in #1432
• fix(provider): use the selected model reserve in the picker by @darkbloombugs in #1438
• fix(provider): stop the start picker at end of input and drain on terminal hangup by @jonathan308 in #1424
• fix(provider): retry unusable startup fan discovery by @jonathan308 in #1529
• Stage SSD checkpoint restores without waiting on another prefill (#867) by @javiech1 in #1519
• fix(provider): advertise forced tools for embedded-MTP Qwen3.8 by @javiech1 in #1527
• fix(ci): recover threat reviews and verify integration against source by @antojoseph in #1533
• feat(nemotron-super): native Nemotron 3.5 Super VL NVFP4 owner with MTP, batching, images and qualification runner by @jonathan308 in #1510
• feat(nemotron-super): serve Nemotron 3.5 Super from a managed production slot (re-opened onto master) by @jonathan308 in #1538
• Serve Nemotron Super video input (mp4/QuickTime) on the managed slot by @jonathan308 in #1528
• fix(provider): reject unsupported Qwen3.8 reasoning effort as a client 400 by @javiech1 in #1523
• perf(nemotron-super): fused MTP verify window and NAX expert prefill on M5 by @jonathan308 in #1531
• GLM-5.3-Flash serving, residency safety and MiMo tool-schema parity by @jonathan308 in #1516
• feat(provider): native Flint support and prefill projection optimization by @jonathan308 in #1535
• fix(ci): recover threat-review failures and preserve scan evidence by @antojoseph in #1542
• fix(gptoss): pin the SDK that keeps GPT-OSS activations and KV in bf16 by @Gajesh2007 in #1539
• perf(native): fused causal SDPA for full-attention prompt chunks by @Gajesh2007 in #1540
• feat(flint): OpenRouter parameter compatibility (reasoning controls, stop, JSON response_format, logprobs) by @jonathan308 in #1552
• fix(install): extract release bundles under umask 022 by @jonathan308 in #1551
• Document the Qwen GDN prompt kernels and forward their rollback to launchd by @Gajesh2007 in #1536
• Stream auto tool calls while they are generated and seed content for plain Qwen prompt tails by @Gajesh2007 in #1549
• perf(provider): chain batch-gated Qwen MTP decode; add concurrent runtime benchmark by @Gajesh2007 in #1543
• perf(provider): reuse prompt tokens across turns and render long Gemma 4 / GPT-OSS histories in linear time by @Gajesh2007 in #1546
• fix(provider): correct the deadline-admission rate estimators by @Gajesh2007 in #1541
• perf(provider): warm first-use kernel shapes before serving by @Gajesh2007 in #1537
• perf(provider): keep busy engines' MLX buffer pool and paged KV segments warm by @Gajesh2007 in #1545
• Raise CBv2 mixed-step prefill geometry for Gemma 4 by @Gajesh2007 in #1547
• fix(flint): serve Flint only on paged attention with default provider config by @jonathan308 in #1554
• perf(flint): pin SDK speedups, chain constrained decode, serve at paged width by @jonathan308 in #1555
• perf(flint): pin exact decode fusions by @jonathan308 in #1556
• Complete automatic threat reviews or hand off to a human by @antojoseph in #1550
• Prepare provider 0.10.0 by @Gajesh2007 in #1557

New Contributors

• @javiech1 made their first contribution in #1520

Full Changelog: v0.9.22...v0.10.0