Changelog

What changed, newest first. Each line links to the guide that documents it.

October 2026

1 to 3 October

Routing aliases fast and best, and fallback chains: send models and the first one that can serve does. Models
Quantization variants (llama-3.1-8b@8bit) and the quality field (economy, standard, max). Models
A long-context tier: requests above 32,768 tokens go to nodes with room for them and bill at 1.5 times. Models
Out-of-range sampler settings are refused with 400 before any engine sees them. Chat completions
Project tags (metadata.project or Silicon-Project) and the cost explorer. Project tags and costs
The OpenAI Files API, so the standard batch flow runs unchanged. Files and batches
LoRA adapters: register one, deploy it to your nodes, call it as base:adp_id. LoRA adapters
Signed receipts countersigned by the gateway, with retired platform keys listed alongside the window they signed in. Verifiable receipts
Prefix-cache affinity: requests that share a long prefix prefer the node that already holds it. Prefix caching
X-Silicon-No-Retention is honored on every engine surface: logs, the event feed and the prefix cache. Privacy
Zero-downtime deploys: a draining gateway answers 503 with Retry-After: 1 and the code draining. Errors

Key protection: a checksummed key format, IP allowlists, anomaly alerts with optional automatic suspension, and GitHub secret scanning. Key protection
Usage that finishes after its key is revoked is still billed, so a request in flight at revocation is not free.
The chat app at /chat: encrypted history with the retention you choose, temporary chats, and a daily spend cap.
The public site in Spanish and German, and model comparison pages.

Top up with USDT or USDC on Arbitrum One, USDC on Base, or USDT on Tron, each credited once the chain finalizes. Billing
Card top-ups through Privy, which buy USDC into your embedded wallet and pay it in. Billing
The wallet binding message now names the site, your organization and the wallet family.
A status page with incidents, 90 days of uptime, and RSS and JSON feeds. Status

Join from the menu bar app or the engine dashboard with device sign-in, and choose the models you offer. Run a node
Model hot swap without dropping in-flight requests, and long prompts prefilled on another node of the same cluster.
An opt-in, signed auto-updater for engines installed as a launch agent, with rollback.
Verification on every surface, earnings held in escrow for a holdback window, and slashing only for failed sample and canary checks. Payouts stay off.
The relay drains gracefully and hands tunnels over to a peer during deploys.
Signed, attested release images and engine builds.

Receipt verification in both SDKs reports bound and honors retired key windows. SDKs
Open WebUI, LiteLLM, n8n and Zapier integrations. Integrations
Streaming parser fixes found by fuzzing, in both SDKs.

1 to 4 September

Embeddings, rerank, speech to text, text to speech and image routes on the gateway.
The Batch API at half price, with a worker that runs batches as their key.
Structured outputs and forced tool calls enforced by grammar-constrained decoding. Structured outputs and tools
Key pinning to regions, operators and attested nodes, and the Silicon-Attestation header.
An honest /health (can inference be served) next to /livez and /readyz.

Teams and roles, invites, two-factor sign-in, Google and GitHub sign-in, and email verification.
Signed webhooks, usage export, deployments that place models on your nodes, and the marketplace.
Cached-token and per-unit pricing.

The relay: nodes behind NAT serve by dialing out, with nothing opened inbound.
The engine gained speculative decoding, continuous batching, Whisper and Kokoro, embeddings and rerank, and a Linux CUDA path.

The TypeScript and Python clients, and the LangChain, LlamaIndex, Vercel AI and Continue integrations.
The project became Silicon Network, at siliconnetwork.io and api.siliconnetwork.io, and the site became a desktop of apps.

Silicon Network›Docs›Guides›Changelog37 items