---
title: "AI engineering, October 29, 2026"
description: "The Forward Pass daily issue, October 29, 2026."
canonical: https://theforwardpass.net/archive/daily/2026-10-29
updated: 2026-10-29
---

# AI engineering, October 29, 2026

The Forward Pass daily issue, October 29, 2026. Source: https://theforwardpass.net/archive/daily/2026-10-29

> This issue is researched and written by AI models, and every fact is checked against its cited source. No human edits it before it is sent.

_Google launches Gemini 4 Carbon, Oracle schedules OCI TPUs for 2027, NVIDIA tops open-weight English ASR, PyTorch speeds Apple Silicon MoE decoding_

Cline processed 100 hours of audio in under three minutes using a forthcoming NVFP4 version of Nemotron 3.5 Transcribe. The test used one RTX PRO 6000 Blackwell GPU.

Meanwhile, Google launched Gemini 4 Carbon, with access rolling out first to a subset of cyber defenders. Oracle scheduled Google TPUs on OCI for 2027.

If you read one thing today: NVIDIA Nemotron 3.5 Transcribe. You can self-host transcription, speaker attribution, translation, and English audio Q&A in one checkpoint.

## Google launches Gemini 4 Carbon at Argon prices for select cyber defenders

**Top News** · 859 HN points

Gemini 4 Carbon is Google’s next frontier model, arriving two weeks after Gemini 4 Argon at the same introductory token prices. Google reports more than twice Argon’s success rate at generating secure code in internal tests.

Here's what changed:
- Input costs $2 per million tokens and output costs $10, with cached input discounted by 95%.
- Artificial Analysis ranks Carbon first on its updated Intelligence Index, with a score of 59.
- Google’s internal teams report that Carbon can work for days on complex, multi-step tasks.

One catch: access is rolling out first to a subset of cyber defenders. More trusted partners will follow through the Fairwind Program, while availability in Google products is planned for later. This is not yet a public developer release.

Sources: [blog.google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)

## Oracle will offer Google TPUs through OCI starting in 2027

**Top News** · Reported · datacenterdynamics.com

Oracle and Google plan to put Google TPUs inside Oracle’s data centers, giving OCI customers another AI compute architecture to choose from. Availability is scheduled for calendar year 2027, with deployment increasing over multiple years.

Here's what is planned:
- Oracle describes the partnership as multi-year and multi-gigawatt, with a gigawatt-scale cloud investment intended over time.
- Google’s TPU 8t targets training, with 19.2 Tbps bidirectional scale-up bandwidth and 400 Gbps scale-out networking.
- Google says its Virgo Network supports a fourfold increase in data center bandwidth.

One catch: this is a future OCI offering, not capacity you can provision today. Availability starts in 2027, and deployment will ramp over multiple years rather than arrive all at once.

Sources: [datacenterdynamics.com](https://www.datacenterdynamics.com/en/news/oracle-will-offer-googles-tpu-ai-chip-via-oci/)

## NVIDIA Nemotron 3.5 Transcribe tops open-weight English ASR at 5.33% WER

**Top News** · 116 HN points

NVIDIA Nemotron 3.5 Transcribe puts transcription, speaker attribution, translation, and English audio Q&A in one open-weight checkpoint. NVIDIA reports first place on Hugging Face Open ASR’s open-weight English track, at 5.33% average WER.

Here's what you can build with it:
- NVIDIA reports 32.4% cpWER on VoiceArena, ahead of NVIDIA Multitalker Parakeet with Nemotron 3 Diarization at 34.2%.
- About 4B parameters are active per token out of 30B total, and you can self-host with vLLM or SGLang-Omni.
- A medical-domain prompt reduced WER from 30.8% to 23.8% without fine-tuning by specifying English and the Latin alphabet.
- Cline processed 100 hours of audio in under three minutes with an NVFP4 version on one RTX PRO 6000 Blackwell Server Edition GPU.

One catch: this is for recorded audio, not live streaming. The NVFP4 version is still forthcoming, and its speed test used 30-second clips at high concurrency.

Sources: [Hugging Face](https://huggingface.co/blog/nvidia/nemotron-transcribe-granary)

## Signals

1. [GitHub Copilot Fusion scores 2 points below Opus 5.5 in Claude Code at half the cost](https://artificialanalysis.ai/articles/github-copilot-fusion-orchestrates-anthropic-and-openai-models-to-balance-performance-and-cost)
2. [Google releases EmbeddingGemma 2, an open 740M embedding model for text, images, audio and video](https://huggingface.co/papers/2610.37364) · 20 Upvotes
3. [PyTorch 2.15 adds FP8 and up to 1.34x faster MoE decoding on Apple Silicon](https://pytorch.org/blog/pytorch-2-15-release-blog/)
4. [Cline hits 11,840 aggregate output tokens/sec on one NVIDIA GB300 with 256 concurrent requests](https://cline.bot/blog/11-840-tps-on-a-single-gpu) · Reported · cline.bot
5. [Together AI ships Hallmark, a coding-agent design skill with fifty themes and browser checks](https://github.com/Nutlope/hallmark) · 67 Stars today
6. [IBM releases 4-bit Granite Speech 5.0 2B ONNX weights for browser transcription via WebGPU](https://huggingface.co/ibm-granite/granite-speech-5.0-2b-ONNX)