--- title: "AI engineering, November 4, 2026" description: "The Forward Pass daily issue, November 4, 2026." canonical: https://theforwardpass.net/archive/daily/2026-11-04 updated: 2026-11-04 --- # AI engineering, November 4, 2026 The Forward Pass daily issue, November 4, 2026. Source: https://theforwardpass.net/archive/daily/2026-11-04 > 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. _Fireworks previews Ultra-fast inference, Revolut cuts coding inference costs roughly 90%, Nvidia tells data centres to reduce China exposure, WebHarbor packages 153 website replicas_ Revolut moved roughly 2,000 engineers onto self-hosted GLM for coding. The fintech reports roughly 90% lower inference costs in less than a year. Meanwhile, Fireworks is previewing Ultra-fast inference on Serverless for DeepSeek V4.1 Flash 1017. Nvidia is directing data centre clients to reduce their exposure to China. If you read one thing today: Fireworks Ultra-fast. Use the dedicated Serverless model ID to test a faster serving mode without changing the underlying model. ## Fireworks previews Ultra-fast, targeting 6-8x generation speed on Serverless **Top News** If token generation is holding up your application, Fireworks has a new serving mode to test. Ultra-fast runs DeepSeek V4.1 Flash 1017 on Serverless without changing the underlying model. You select it with a dedicated model ID rather than moving to a different model. Here's what changed: - Ultra-fast variants aim for 6-8x generation speed. - Agentic coding targets go as high as 10x, according to Fireworks. - Output quality stays the same because Ultra-fast changes the inference mode, not the model. One catch: Ultra-fast is in Preview and under active development, so its behavior or availability may change quickly. Try it: On Fireworks Serverless, set the model to `accounts/fireworks/routers/deepseek-v4p1-flash-1017-ultrafast`. Sources: [docs.fireworks.ai](https://docs.fireworks.ai/updates/changelog) ## Revolut reports roughly 90% lower AI coding inference costs with self-hosted GLM **Top News** Revolut made self-hosted GLM its default for AI coding, reporting roughly 90% lower inference costs than relying primarily on commercial services such as Claude. This is an internal deployment for its engineers, not a public coding service. Here's what changed: - Revolut reported GLM-5.1 inference at $0.30 per million input/output tokens, about one-tenth its stated cost for OpenAI Codex or Anthropic Claude Code. - Serving optimizations brought time to first token below three seconds at around 256,000 tokens of context, with generation around 200 tokens per second, Revolut reported. - The rollout reached roughly 2,000 engineers. Revolut says coding time improved 60% and lead time improved 70%. - Revolut moved from GLM-5.1 to GLM-5.3, serving GLM-5.3 on 56 Nvidia B200 GPUs. One catch: GLM did not fully replace GPT or Claude. It is text-only and not ideal for UI work, so Revolut routed vision queries to other models. Sources: [theregister.com](https://www.theregister.com/ai-and-ml/2026/11/03/british-fintech-revolut-slashes-ai-coding-costs-10x-by-cutting-out-claude-and-going-self-hosted/5307345) ## Nvidia tells cloud providers most clients must be non-Chinese **Top News** · Reported · ft.com Nvidia is putting a customer-base requirement on cloud computing providers. It is directing data centre clients to reduce their exposure to China, and says a majority of cloud providers' clients must be non-Chinese. Here's what the direction covers: - Data centre clients are being told to reduce their China exposure. - Cloud computing providers have a specific instruction: their client base must have a non-Chinese majority. - The majority requirement concerns the provider's customers, rather than the provider's own nationality. Why care? If you are discussing this directive with a cloud provider, the concrete question is about who that provider serves. Nvidia told cloud computing providers that a majority of their clients must be non-Chinese, while directing data centre clients to reduce their exposure to China. Sources: [ft.com](https://www.ft.com/content/1902df70-1380-4588-bad5-171352b575f2?syn-25a6b1a6=1) ## Signals 1. [Axion 1.0 uses three training stages for an open agent foundation model](https://huggingface.co/papers/2611.03233) · 20 Upvotes 2. [WebHarbor packages 153 real-world website replicas into Docker environments for GUI agents](https://huggingface.co/papers/2611.03634) · 2 Upvotes 3. [Mamba, Gated DeltaNet and xLSTM risk training stalls from near-zero token embeddings](https://huggingface.co/papers/2611.04449) · 6 Upvotes 4. [DR Pixel runs deep research over 1.06M document pages using images, not OCR](https://huggingface.co/papers/2611.03097) · 5 Upvotes 5. [Scrapling handles web scraping from single requests to full-scale crawls](https://github.com/D4Vinci/Scrapling) · 90 Stars today 6. [Krea 3 finishes image edits within 8 seconds in production](https://www.krea.ai/blog/krea-3-technical-report)