NIBIA Fabric v0.7.0 Experimental Alpha

NIBIA Fabric v0.7.0 Experimental Alpha

NIBIA Fabric v0.7.0-alpha is the first public Experimental Alpha of NIBIA's CPU-first local AI compute fabric.

A Primary computer can combine its own local compute/memory with safe capacity from other computers on the same trusted LAN so larger GGUF models can run across heterogeneous everyday hardware.

Release identity

• Public CLI version: v0.7.0-alpha
• NIBIA wire protocol: 2
• Managed llama.cpp: b10902, commit df03399
• License: Apache-2.0
• The data-plane stabilization baseline originated in the alpha.6.18.7 line. Pre-public candidates then narrowed the release surface and hardened CLI, lifecycle diagnostics, installers, and documentation without changing wire protocol 2 or the pinned llama.cpp runtime.

Highlights

• N-node heterogeneous CPU and RAM capacity expansion.
• macOS Apple Silicon, Linux amd64, and Windows amd64 nodes; any supported platform can be the Primary Node.
• Minimum Safe Fabric / Lazy Worker Activation.
• Adaptive SAFE memory reserve with visible SAFE USABLE capacity.
• Linux/Windows CPU Primary execution and Apple Silicon Metal local execution.
• On-demand llama.cpp RPC workers carried through authenticated reverse mTLS relays.
• Raw llama.cpp RPC remains loopback-only on Workers.
• Pinned and SHA-256-verified precompiled llama.cpp runtime.
• Persistent worker-local RPC tensor cache.
• Workload-scoped Power Guard on macOS, Linux, and Windows.
• Local Web UI plus OpenAI-compatible API.
• Automatic bounded SERVE service recovery after selected-worker loss and reconnection.
• Prebuilt per-platform archives and user-local installers, including checksum-first macOS quarantine guidance for an unsigned Experimental Alpha.
• Focused user CLI with top-level nibia run for one-shot/terminal inference and nibia serve for persistent Web UI/API serving; lower-level Fabric diagnostics live under nibia fabric status/devices/worker/relay.
• Consistent -h / --help behavior across user and advanced command namespaces, with side-effect-free help and documented --flag rendering.
• Simplified nibia pair pairing-code creation and explicit nibia run prompt semantics: one-shot RUN requires --prompt; interactive --session may start without one.
• Consistent RUN/SERVE staged UX with shared safe-capacity and tensor-cache/reuse reporting. Cache-aware distributed loading reports observed relay transfer, rate, and elapsed time rather than a misleading percentage/ETA against planned remote tensor placement. --verbose exposes detailed NIBIA planning/telemetry; --verbose-runtime remains the raw llama.cpp debug surface.
• RUN and SERVE use the same live cache-aware loading cadence. When an advanced relay is stopped, relay status distinguishes the still-ready Agent reverse-mTLS tunnel pool from the inactive controller-loopback relay and retains the node name.
• Managed RPC worker diagnostics distinguish requested cleanup from unexpected exits, so an intentional worker stop is not reported as Last exit error.
• RUN/SERVE lifecycle cleanup waits for the Controller/Agent cleanup jobs instead of applying a stricter CLI heartbeat-staleness cutoff; Runtime cleanup: complete is reserved for warning-free teardown.
• One-shot/terminal RUN and advanced device discovery now tear down transient managed RPC workers and authenticated relays on command completion or failure instead of leaving diagnostic/workload processes behind.

Final exact-artifact acceptance

The exact public-named macOS arm64, Linux amd64, and Windows amd64 archives were reinstalled on the reference hardware before publication. nibia version / protocol 2, local Doctor, installed-binary identity, three-node Fabric parity, distributed RUN, localhost-only SERVE/API, and synchronous lifecycle cleanup all passed on the frozen release bytes.

Frozen release-asset SHA-256 values:

• 661a02243a946e0539ad7eb202b43d157284e3078389ec6252b15e80503fbb17 — nibia-v0.7.0-alpha-macos-arm64.zip
• d6622c93c597f3422d89d836ed712a60578a16d54a731b9ad28633583c706038 — nibia-v0.7.0-alpha-linux-amd64.zip
• f166cb3078f035ad8e46a2209ea1a266e519b8c194368144a83a192e3e35c898 — nibia-v0.7.0-alpha-windows-amd64.zip

Publication note (2026-09-27): the checksum values in the release notes were corrected after publication to match the already-published release assets and SHA-256 manifest. The release assets and tag were not replaced or modified.

The macOS archive remains an unsigned Experimental Alpha; the installer therefore requires checksum verification before documented quarantine removal when Gatekeeper quarantine is present.

Physical validation matrix

The release was validated on a heterogeneous three-machine fabric consisting of Apple Silicon macOS, amd64 Linux, and amd64 Windows nodes over a local LAN.

Qwen family

• Qwen3 4B Q4_K_M — small-model/local and distributed execution coverage during development.
• Qwen3 14B Q6_K — successful heterogeneous distributed inference, including three-node execution.
• Qwen3 30B-A3B Q4_K_M — primary release-validation workload. Validated for capacity expansion, repeated lifecycle/cleanup, selected-worker fault handling, automatic service recovery, idle/load stability, OpenAI-compatible serving, and contexts of 8192, 16384, and 32768 when the reference fabric had sufficient SAFE memory. Open WebUI and OpenClaw client validation is listed separately below.

Cross-family release smoke

• GPT-OSS 20B MXFP4 — capacity-expanding distributed load and real OpenAI-compatible inference.
• Gemma 3 27B IT Q4_K_M — capacity-expanding dense-model distributed load and real OpenAI-compatible inference.

Other llama.cpp-compatible GGUF models may work but are not necessarily validated in this release.

Reference benchmark highlights

• Qwen3-30B-A3B / three-node Gigabit Ethernet: 17.28 GiB model; 9.38 GiB first-load transfer at 76.4 MiB/s; model ready in 2:06; 9.337 tok/s median across three no-thinking generation runs.
• Qwen3-14B Q6_K / Mac + Windows A/B: model ready 65 s on Wi-Fi versus 27 s on Gigabit Ethernet; median generation throughput 3.47 versus 4.57 tok/s respectively in the measured topology.
• Warm-cache stabilization acceptance: Qwen3-30B-A3B planned 10.80 GiB of remote tensor payload while only 0.32 GiB of relay traffic was observed, avoiding an estimated ~10.48 GiB of repeated transfer. The residual cache-miss path remains a known post-alpha optimization target.

See docs/BENCHMARKING.md for hardware, raw run values, methodology, and interpretation caveats.

Client compatibility validated

• llama.cpp built-in/local Web UI — the default interactive UI served by NIBIA.
• OpenAI-compatible /v1/models.
• OpenAI-compatible /v1/chat/completions.
• Open WebUI — physically validated against NIBIA Fabric's local OpenAI-compatible API, including release tests at 8192 and 16384 context sizes.
• OpenClaw — physically validated in local mode with NIBIA Fabric configured as an OpenAI-compatible provider at an 8192-token context window.

Open WebUI and OpenClaw are optional independent clients, not required NIBIA dependencies. Other OpenAI-compatible clients may work but are not necessarily validated in this release.

The OpenAI-compatible endpoint remains local to the Primary by default: http://127.0.0.1:8081/v1.

Recovery semantics

If a selected distributed worker disappears during an active generation, the active request is interrupted rather than continuing with an unsafe partial fabric.

By default, SERVE remains alive for a bounded recovery window, waits for the selected capacity to return, rebuilds the remote RPC fabric, reloads the model, and restores the API automatically when possible.

This release does not implement transparent request continuation, request replay, distributed KV-cache replication, or migration of an in-flight generation.

Known limitations

• Experimental Alpha: CLI/API/state details may change.
• Release scope: this alpha ships the distributed GGUF inference path. RAG, embedding pipelines, Ollama orchestration, and generic synthetic distributed workloads are not part of this release.
• Trusted LAN only: pairing bootstrap is not designed for hostile/public networks.
• Local-only SERVE: the first alpha intentionally rejects non-loopback inference binding.
• CPU-first: GPU pooling is not part of this release.
• Inference only: LoRA/QLoRA training is not part of this alpha.
• Power policy: Power Guard cannot override every critical-battery, lid-close, firmware, manual-sleep, or administrator policy.
• Model provenance: managed llama.cpp downloads are verified; arbitrary GGUF model files are not signed/verified by NIBIA.
• Capacity-first scheduler: Minimum Safe Fabric prioritizes safe placement; more nodes do not necessarily mean higher tokens/s.
• Tensor-cache performance: persistent RPC tensor reuse is functional, but some warm loads with small residual misses can still spend disproportionate time in cache resolution/relay work. This is a post-alpha performance optimization item, not a correctness blocker.
• macOS signing: if the published Experimental Alpha archive is not yet Developer ID signed/notarized, verify the SHA-256 first and follow the documented quarantine-removal step only for the verified archive.

Networking guidance

Ethernet is recommended for larger distributed models when practical. Wi-Fi is supported. VPNs can block or reroute LAN traffic; enable local-network access or disable the VPN when troubleshooting fabric reachability. On multi-interface nodes, --preferred-address <NODE_LAN_IP> can pin the Agent's advertised Fabric address to the intended physical LAN interface.

Power guidance

For long workloads, keep laptops connected to external power when practical and configure them not to auto-sleep. NIBIA Power Guard is an additional workload-scoped safeguard, not a replacement for host power policy.