vLLM Updates 2026

Three stand out. Semantic Router v0.1 (codename Iris) shipped on January 5, 2026 as the first major release for intelligent cross-model routing, landing with over 600 merged pull requests and 50+ contributors since its September 2025 experimental launch. v0.18.0 followed in late March 2026, adding gRPC serving via the new --grpc flag, GPU-accelerated NGram speculative decoding compatible with the async scheduler, and the vllm launch render command for GPU-less multimodal preprocessing. v0.19.0 landed on April 3, 2026 with Day 0 Gemma 4 support (including first-ever Day 0 support on Google TPUs), Model Runner V2 maturation, zero-bubble async scheduling compatible with speculative decoding, CPU KV cache offloading, a new /v1/chat/completions/batch endpoint, and NVIDIA B300/GB300 support.

Iris routes inbound requests to different backends based on request shape, cost, and intent, as an inference-server-level primitive rather than something applications hand-roll. It is the inflection point of 2026 because it reframes vLLM from a single-model serving engine into a routing fabric. The same pattern has appeared in client-side agents. Fazm's ACP bridge dispatches each chat to a selected model via a session/set_model RPC call at acp-bridge/src/index.ts line 1332, with a runtime-updated list of available models emitted by emitModelsIfChanged at line 1132. Server-side routing in vLLM and session-level routing on the client are two sides of the same 2026 shift, and they compose: you can route in Fazm to a local vLLM endpoint that itself routes via Iris.

Fazm's macos-use tool reads Terminal.app's accessibility tree directly as structured text, including cursor position, scrollback, and visible lines. The capture entry point is in Desktop/Sources/AppState.swift around line 439, using AXUIElementCreateApplication against the frontmost process. The tree arrives as lines of the form [AXStaticText] "INFO 04-03 12:14:02 engine.py:321] Model loaded in 18.4s" x:72 y:312 w:640 h:16. A text-only LLM can parse that verbatim, without a vision encoder, and without OCR error on dense vLLM log output. Typical captures fall in the 1 KB to 42 KB range per turn, which is what keeps context usage sane across a long upgrade watch.

Indirectly, yes. Fazm does not ship a vLLM integration per se, but it exposes a single setting that makes any compatible endpoint addressable. The UI lives at Desktop/Sources/MainWindow/Pages/SettingsPage.swift (Custom API Endpoint field, under Settings > Advanced). At session start, Desktop/Sources/Chat/ACPBridge.swift lines 378 to 380 read that value and set env["ANTHROPIC_BASE_URL"] on the ACP subprocess before spawn. If you run a shim that translates Anthropic Messages API requests into vLLM's /v1/chat/completions shape, Fazm will talk to your vLLM deployment for every tool-using chat, using the accessibility tree as primary context instead of screenshots.

Three things. First, gRPC serving via the --grpc flag, which is faster for fan-out clients that currently spin up an HTTP session per request. Second, GPU-accelerated NGram speculative decoding, now compatible with the async scheduler, which makes spec decode a net win rather than a wash on many workloads. Third, the vllm launch render subcommand, which decouples multimodal preprocessing (image tokenization, resizing, feature extraction) from GPU inference so a small pool of CPU workers can feed a single expensive GPU. The operational consequence: your old launch scripts still work, but the two-process render/serve split is the new recommended shape if you run any multimodal models.

Day 0 means the weights run on vLLM the day Google publishes them. Day 0 on TPUs specifically means you can get Gemma 4 onto Cloud TPU pods without writing a separate serving stack. For teams running a mixed fleet (some H100s, some TPUs, some B300s), that collapses a fragmentation problem. For solo practitioners, it raises the floor: vLLM is now a serving target for the three major accelerator families, not an NVIDIA-first project with bolted-on other backends.

Because it is not the engine's story, and the release notes write themselves. 'What shipped' is easy copy: read GitHub releases, quote the blog, link the roadmap. 'How an operator runs the upgrade' needs primary reporting: which dashboards do they watch, what do their benchmark scripts look like, what breaks between versions. On the Mac side specifically, the friction is that SSH + browser dashboards + a local editor + a spreadsheet for the comparison is a cross-app workflow, not a single-app one. A screenshot-based agent cannot read that reliably because the screenshot-to-text loop slips on dense log output and loses context fast. A text-tree agent (Fazm) can.

Yes. Three files and line ranges. Default model: open acp-bridge/src/index.ts and look at line 1126 for const DEFAULT_MODEL = "claude-sonnet-4-6". Model routing: line 1332 in the same file, await acpRequest("session/set_model", { sessionId, modelId: requestedModel }). Endpoint redirection: Desktop/Sources/Chat/ACPBridge.swift lines 378 to 380, env["ANTHROPIC_BASE_URL"] = customEndpoint when the Custom API Endpoint setting is filled. Pair that with your vLLM v0.19.0 server (or an Anthropic-to-OpenAI shim in front of it) and Fazm routes every chat to vLLM instead of Anthropic's servers.