# Cerebras disaggregated inference, Scribe v3 Live Medical, Slurm HRES

> Cerebras splits prefill and decode across chips, ElevenLabs tunes live transcription for clinical calls, and Slurm 26.11 learns to schedule QPUs.

Published 18 Nov 2026. Coverage: 17 Nov 2026 to 18 Nov 2026. By Ali Asif.

## Items

1. **Cerebras launches disaggregated inference** (Cerebras, Frontier): Prefill runs on GPUs and decode on Cerebras CS-3, with a reported 5x throughput per megawatt on a GPT-OSS-120B endpoint versus a CS-3-only setup. [Primary source](https://www.cerebras.ai/blog/cerebras-launches-disaggregated-inference)
2. **Scribe v3 Live Medical** (ElevenLabs, Agents and coding): A live speech-to-text model tuned for clinical language, reporting 12.8% WER on PriMock57 and custom term lists of up to 10,000 terms. [Primary source](https://elevenlabs.io/blog/scribe-v3-live-medical)
3. **Integrating QPUs into Supercomputing Centers Using Slurm HRES Plugins** (NVIDIA, Science): Slurm 26.11 adds Hierarchical Resources plugins so a QPU can be scheduled as a cluster-level resource instead of being tied to one node. [Primary source](https://developer.nvidia.com/blog/integrating-qpus-into-supercomputing-centers-using-slurm-hres-plugins)

Three items today, all from pages I opened and read. I have not run any of it, so every number below is the vendor's own. OpenAI also posted about GPT-6.2 Astra and Sol, but its page sat behind a bot check for me, so I am leaving it out until I can read it directly.

## Cerebras splits prefill and decode

Cerebras says disaggregated inference is now in production. Each request is divided in two. Prompt processing, called prefill, runs on GPUs such as AMD Instinct. Response generation, called decode, runs on Cerebras CS-3 systems. The reasoning is that decode is autoregressive and limited by memory bandwidth, and Cerebras points to its large on-chip SRAM, claiming 350x the SRAM capacity and 5,400x the memory bandwidth of a typical GB300 GPU.

The measured result is about 5x throughput per megawatt compared with a CS-3-only configuration, on a GPT-OSS-120B endpoint on November 14, with no loss in token generation speed. The post also shows a roughly 10x decode advantage, but that is an illustration, not a measurement.

For partners, there are two routes. One is an end-to-end stack with a router, a vLLM or SGLang prefill engine, a Cerebras decode engine and a KV Portal that moves model state over RDMA. The other is a decode-only SDK for teams that bring their own prefill. Amazon Web Services, Fireworks AI, Gimlet Labs and General Compute are early-access partners, and Cerebras plans hundreds of megawatts of this by the end of 2027.

## ElevenLabs Scribe v3 Live Medical

This is a real-time transcription model built on Scribe v3 Live and tuned for clinical conversations, available in ElevenAgents. ElevenLabs reports 12.8% WER on PriMock57, the lowest among the medical models it tested and about 17% fewer errors than the next closest. On Fareez OSCE it reports 0.6% semantic WER. Everyday speech is roughly unchanged from Scribe v3 Live, and latency stays around 150 ms.

The practical feature is custom term lists of up to 10,000 terms, which the company says cut errors on medical terms by up to 48%. It also waits for a caller to finish reading out a member ID before the agent replies. The caveat is in the post: ElevenLabs ran the benchmarks itself, on selected providers in real-time mode.

## Slurm learns to schedule QPUs

NVIDIA's post covers Slurm 26.11, which adds Hierarchical Resources, or HRES, plugins. The old approaches had gaps: GRES pins a device to one node, so a node failure strands the QPU, and the license-plus-SPANK workaround can leave compute nodes idle while a job waits. With HRES, a site describes the QPU in a resources.yaml file and users request it with `--resources=qpu:1`. The first implementation is hres/lua.

Plugin callbacks handle health, admission, reservation and release, and a remote failure keeps the job pending instead of draining the node. The community QRMI and QDMI projects have released plugins on these APIs, and Alice & Bob, IBM, IQM, MQSC, Pasqal and Quantinuum are listed as adopters.

## What I'd watch next

Independent numbers on Cerebras disaggregation, ideally from one of the early-access partners. Whether Scribe Medical holds its WER on accents and noisy lines outside ElevenLabs' own test sets. And which Slurm sites turn on HRES for real QPUs first.

## My take

The Cerebras result is the one I would check first, because the 5x figure comes from one endpoint on one day and the 10x decode picture in the post is an illustration, not a measurement. Scribe Medical matters to me for the term lists, since drug names are where voice agents fail in practice. The Slurm post is niche, but it shows how plain scheduler plumbing decides whether quantum hardware gets used at all.

## Sources

- https://www.cerebras.ai/blog/cerebras-launches-disaggregated-inference
- https://elevenlabs.io/blog/scribe-v3-live-medical
- https://developer.nvidia.com/blog/integrating-qpus-into-supercomputing-centers-using-slurm-hres-plugins

Canonical: https://heyaliasif.com/ai/2026-11-18-cerebras-disaggregated-scribe-medical-slurm-hres