Darkbloom — The AI Toolchain

Darkbloom v0.8.15

Sat, 29 Aug 2026 02:11:17 GMT

Darkbloom v0.8.15 adds opt-in JSON metadata on chat completions and integrates the Qwen3.8-27B VLM. • Adds `metadata_details: true` request body field to `POST /v1/chat/completions` to embed provider attestation, timing, and job identity details directly in the JSON response body — useful when OpenAI-compatible SDKs hide raw HTTP headers. • Adds `X-Darkbloom-Metadata-Details: true` request header as an alternative opt-in for the same per-request metadata embedding, supporting browser preflight flows. • When metadata is opted in, `metadata.location` surfaces the serving provider's GeoIP region as `region`, `region_code`, `country`, `country_code`, and `timezone` fields; city, coordinates, and raw IPs are intentionally omitted. • Integrates Qwen3.8-27B full VLM (`EigenLabs/Qwen3.8-27B-4bit`) with vision and video (API) support, restricted to Apple M5 hardware with the approved NAX runtime. • Adds `EigenLabs/Qwen3.8-27B-MTP-4bit` as an MTP proposal assistant for Qwen3.8, enabled by default; disable with `mtp_mode = "off"`, `darkbloom beta disable mtp`, or the `DARKBLOOM_CBV2_MTP=0` environment variable.

Darkbloom v0.8.14

Thu, 27 Aug 2026 22:09:25 GMT

Darkbloom v0.8.14 adds Qwen3-VL production serving, automatic Qwen MTP defaulting, and competing-inference warnings in `darkbloom doctor`. • Adds `mtp_mode` config key accepting `'auto' | 'on' | 'off'` to control Multi-Token Prediction; valid inline Qwen3.5/3.6 MTP artifacts now default to `auto` (on) while Gemma remains opt-in; `DARKBLOOM_CBV2_MTP=0` env var remains an independent rollback path. • Config schema v3 migrates legacy generated `mtp = false` values to `auto` and retains legacy `mtp = true` as `on`, so upgraded providers receive the new policy automatically. • Adds `competing inference` check to `darkbloom doctor` that warns when local inference processes (e.g. Ollama on port 11434, llama-server) are detected consuming unified memory alongside the provider. • Adds production serving of Qwen3-VL (`qwen3_vl_moe` architecture) through the ContinuousBatchingV2 path, with per-row M-RoPE for text, causal visual spans and every DeepStack level for image prefill, and fused homogeneous routed gate/up expert projections at load time.

Darkbloom v0.8.13

Wed, 26 Aug 2026 06:40:49 GMT

Darkbloom v0.8.13 adds Qwen 3.5 video inference, grounded image captions, and hardware serial number redaction from public APIs. • Supports Qwen 3.5 video inference via EngineV2, sampling at most 8 frames at no more than 512×512 pixels per frame and bounding the vision score tensor to 512 MiB, replacing the previous `400 invalid media input` failure. • Adds EXIF-orientation-aware full-raster image decode for Qwen vision requests, defaulting `enable_thinking=false` for media unless the client explicitly sets `reasoning.enabled`. • Removes hardware serial numbers, UDIDs, and raw MDA certificates from public APIs, headers, URLs, and enrollment flows; public surfaces now expose only opaque provider IDs and trust-verdict status. • Updates earnings calculators to require Mac model, chip family, and unified memory inputs before showing an estimate, grounded in bandwidth-limited capacity; adds M5 Ultra and M6 hardware profiles. • Pins `libs/mlx-swift-lm` to commit `fe01df9`, adding Qwen3-VL 30B-A3B model support.

Darkbloom v0.8.12

Tue, 25 Aug 2026 18:00:16 GMT

Darkbloom v0.8.12 enables atomic first-token deadline admission by default, with typed TOML override support. • Sets atomic first-token deadline admission to default-on for eligible requests, returning a health-neutral `deadline_unreachable` before GPU prompt work begins so the coordinator can retry another provider.

Darkbloom v0.8.9

Fri, 21 Aug 2026 23:33:14 GMT

Darkbloom v0.8.9 adds concrete-source support for OpenRouter aliases, exposing them via a dedicated `/v1/models/openrouter` endpoint. • Adds support for OpenRouter-only aliases to reference concrete catalog models as their source, not just standard rollout aliases; alias source kind is persisted to prevent silent retargeting on later takeovers. • Exposes OpenRouter-only aliases exclusively through `/v1/models/openrouter`, omitting them from the main `/v1/models` list while preserving exact `GET /v1/models/{id}` retrieval — including at zero connected providers. • Shares source pricing, limits, features, readiness, capacity, and datacenter data in the dedicated OpenRouter alias feed.

Darkbloom v0.8.7

Thu, 20 Aug 2026 20:25:00 GMT

Darkbloom v0.8.7 exposes OpenRouter aliases in the model catalog via `GET /v1/models`. • Surfaces OpenRouter-only model aliases in `GET /v1/models`, making previously undiscoverable model IDs available for direct inference requests.

Darkbloom v0.8.6

Thu, 20 Aug 2026 17:03:36 GMT

Darkbloom v0.8.6 ships default-on CBv2 prefill stack (+38% prefill speed), OpenRouter model aliases, and new MTP metrics. • Adds `DARKBLOOM_CBV2_SOLO_PREFILL_STRIPE=0` environment variable to disable the new solo-prefill stripe (2048-token chunks), which regresses TTFT ~12% under Low Power Mode — throttled or battery-backed providers should set this escape. • Adds `DARKBLOOM_CBV2_PREFILL_NARROWING=0` environment variable to restore pre-v0.8.6 behavior where intermediate chunks return the full `[1,512,248320]` logits tensor (242.5 MiB/chunk) instead of a one-element handle. • Adds `DARKBLOOM_CBV2_MAX_PARTIAL_PREFILLS=1` opt-in environment variable to cap prompt-work rows per scheduler step (FCFS), turning burst TTFTs into a staircase instead of all requests waiting for the full makespan. • Enables CBv2 prefill stack by default, delivering ~1,766 tok/s at 8K on M4 Max (+38% vs v0.8.5 defaults) via expert-tile trust serving default, solo-prefill stripe, recurrent prompt narrowing, and packed prefill for Qwen3.6. • Adds OpenRouter-only model aliases in the coordinator, expanding the set of model IDs routable through OpenRouter. • Promotes adaptive persistent-history MTP with persistent head KV and adaptive width selection onto the release pin (still behind the `mtp` beta flag).

Darkbloom v0.8.5

Sat, 15 Aug 2026 00:40:38 GMT

Darkbloom v0.8.5 adds MTP posture and acceptance counters to the local `/metrics` endpoint and human-approved agent production access. • Exposes `mtp_enabled`, `mtp_active`, `mtp_rounds_total`, `mtp_tokens_proposed_total`, `mtp_tokens_accepted_total`, and `mtp_inactive_reason{model,reason}` counters on the local `/metrics` endpoint in both `--local` and unified serving modes, making MTP acceptance observable without Datadog Logs. • Allows human-approved agents to operate in production environments. • Adds Qwen3.6 E=256 expert-tile prefill route with fused gate/up projection (SwitchGLU(fuseGateUp: true)) and an opt-in `trust` refinement, delivering up to +15.2% prefill throughput at 8k context on M4 Max. • Adds adaptive persistent-history capture-verify MTP stack for Qwen3.6 (behind the `mtp` beta flag, default off), measured at 2.0077x throughput vs non-MTP in canary validation.

Darkbloom v0.8.3

Thu, 13 Aug 2026 07:15:29 GMT

Darkbloom v0.8.3 adds Qwen3.6-35B-A3B VLM with inline MTP on the production provider path. • Adds production-path text, image, and tool inference for the combined Qwen3.6-35B-A3B VLM artifact via the EngineV2 provider path, with request-owned recurrent state, three-axis mRoPE state, causal vision attention, exact rollback, and source-matched target/assistant memory accounting. • Ships inline MTP (depth-one, serial, exact-target-verified) alongside the Qwen3.6-35B-A3B model; video, prefix reuse, paged KV, compiled decode, packed prefill, and rectangular MTP remain fail-closed.

Darkbloom v0.8.2

Wed, 12 Aug 2026 02:01:14 GMT

Darkbloom v0.8.2 adds self-serve payout bank updates for providers, Gemma 4 26B config-controlled optimizations, and tightened provider privacy controls. • Enables `[gemma_optimizations]` config controls for Gemma 4 26B-A4B v0.8.2, gating layer-18 lazy prefill submission, coupled weighted-expert-unsort, and safe-R1 expert-QMM gate (both default-on). • Providers can now update their own Stripe Express payout bank account via a new self-serve endpoint, removing the need to contact support for bank account changes. • Keeps inline video (MP4/QuickTime) off disk by decoding coordinator-inlined bytes through a bounded, memory-backed AVFoundation asset and purging any legacy `vlm-.mp4` files on launch. • Closes provider-derived plaintext egress paths: inference failures cross the WebSocket boundary only as closed-vocabulary codes/reasons; free-form telemetry and automatic provider log reporting are retired in favor of the operator-initiated `darkbloom report` command (with `--dry-run` support and authenticated upload).

Darkbloom v0.8.0

Wed, 29 Jul 2026 20:33:30 GMT

Darkbloom v0.8.0 completes PagedAttention migration and adds live-togglable remote media fetching at the coordinator. • Adds `EIGENINFERENCE_MEDIA_FETCH_ENABLED=false` environment variable to disable remote media fetching live — no redeploy required — restoring the prior HTTP 400 response for requests with remote image or video URLs. • Accepts OpenAI-style remote `http(s)` `image_url` and `video_url` parts on `POST /v1/chat/completions`: the coordinator fetches each URL once, structurally validates it, and rewrites it as an inline base64 `data:` URI before coordinator-to-provider encryption — keeping provider Macs from ever fetching request-controlled URLs. • Rejects remote media URLs in sender-sealed requests with instructions to inline them, preserving the end-to-end encryption guarantee for sealed payloads.