# Anthropic's Scion, a 4-step Qwen-Image LoRA, cuDF skills

> Anthropic replaces git for its monorepo, Qwen-Image-2.1 gets a 4-step LoRA, NVIDIA ships agent skills for Spark UDFs, Gemma detects scams on-device.

Published 14 Oct 2026. Coverage: 13 Oct 2026 to 14 Oct 2026. By Ali Asif.

## Items

1. **Redesigning source control at Anthropic** (Anthropic, Agents and coding): Anthropic built Scion, its own version control system, and moved its monorepo off git in five months as merges grew from about 2,500 a day in May to over 30,000 in September. [Primary source](https://www.anthropic.com/engineering/redesigning-source-control)
2. **Qwen-Image-2.1-Lightning** (LightX2V, Open weights): A 4-step distilled LoRA for Qwen-Image-2.1 covering text-to-image and editing, trained on 16 Huawei Ascend 910B NPUs. [Primary source](https://huggingface.co/lightx2v/Qwen-Image-2.1-Lightning)
3. **How to Rewrite Apache Spark UDFs for GPU Acceleration with AI Agents** (NVIDIA, Agents and coding): cuDF skills for Apache Spark guide a coding agent through rewriting UDFs; one UDF went from 274.55 to 12.78 operator hours on 600 billion rows. [Primary source](https://developer.nvidia.com/blog/rewrite-apache-spark-udfs-gpu-acceleration-ai-agents)
4. **Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection** (Google DeepMind, Policy and business): A fine-tuned Gemma model with a lightweight first-stage classifier runs on the device, with rollout planned in phases and no performance figures published yet. [Primary source](https://deepmind.google/blog/on-device-ai-scam-detection-mddi)

A quieter day after the Gemini 4 Argon launch, but three of today's items are about how work gets built and run once agents do a lot of it. I only used pages I opened.

## Anthropic replaces git for its monorepo

Anthropic wrote up Scion, an in-house version control system. Merges per day grew from about 2,500 in May 2026 to over 30,000 in September, main branch history reached almost 100 GB, and CI ran around 25 million checkouts a day. The team migrated off git in five months.

The design starts from the fact that clients are almost always online. The server handles history, blame, commits, merges and rebases, the client keeps a bounded cache, and the working copy is optional and sparse. Anthropic reports a CPU and memory footprint about 200x smaller than the old git mirroring setup, checkouts roughly two orders of magnitude faster, a 25% CPU capacity gain in CI, and cache hit rates above 99% in every tier. They also say plainly that Scion is not a general replacement for git.

One detail I liked: Claude rewrote the git-to-Scion ingestion pipeline, reportedly with 20x less memory and 4x faster, while engineers kept the design intent and said no when it broke system constraints.

## A 4-step LoRA for Qwen-Image-2.1

LightX2V released Qwen-Image-2.1-Lightning, a distilled LoRA that runs the base model in 4 steps instead of the 40 shown in the comparison on the page, with CFG disabled. It covers text-to-image and image editing with one or more reference images. Training ran entirely on 16 Huawei Ascend 910B NPUs using DMD in LightX2V. The model page does not state a license, so check before commercial use.

## NVIDIA's cuDF skills for Spark UDFs

NVIDIA describes skills for coding agents that rewrite CPU-bound Spark UDFs for the GPU through SQL, cuDF or custom CUDA kernels, with test, convert, benchmark and optimize stages. The headline case is one UDF touching more than 3,000 jobs, cut from 274.55 to 12.78 operator hours on 600 billion rows, a 22x speedup. A large consumer technology company converted more than 30 production UDFs this way, per the post.

## Gemma on the device for scam detection

Google DeepMind and GovTech Singapore built a two-stage scam detector: a low-power classifier screens messages, then a fine-tuned Gemma model reads more of the conversation and warns the user. Training used anonymised scam report data, and neither side can access user conversations. No accuracy figures are published yet, and rollout will be phased.

## What I'd watch next

Whether other large engineering teams publish their own agent-era source control numbers, whether the Lightning LoRA gets a clear license, and whether DeepMind shares detection results for the scam model.

## My take

The Anthropic post is the one I would read twice. When the clients are agents and CI jobs that are always online, a git design built for offline humans stops fitting, and the numbers they report show it. The NVIDIA piece makes the same point from another side: agents do the rewriting, the checks decide what ships.

## Sources

- https://www.anthropic.com/engineering/redesigning-source-control
- https://huggingface.co/lightx2v/Qwen-Image-2.1-Lightning
- https://developer.nvidia.com/blog/rewrite-apache-spark-udfs-gpu-acceleration-ai-agents
- https://deepmind.google/blog/on-device-ai-scam-detection-mddi

Canonical: https://heyaliasif.com/ai/2026-10-14-scion-qwen-lightning-cudf-skills