SmallCode: A Coding Agent That Hits 87% on Benchmarks with a 4B-Parameter Local Model

A developer known as Glittering_Focus1538 has announced a coding agent called "SmallCode" on Reddit's LocalLLaMA community, and it's drawing a huge response.

Existing coding agents like Codex and Claude Code assume the use of frontier models such as GPT-5.4 or Claude Opus. As a result, running them with local models like Gemma or Qwen often leads to failed tool calls, context overflow, and the collapse of multi-step tasks.

SmallCode tackles this problem head-on. It uses Gemma 4 (activating only 4B parameters per token) and scored 87/100 on benchmark tasks — beating the roughly 75% achieved by OpenCode, which relies on a 14B model.

The key techniques are as follows:

Composite tools**: The four steps of file search → read → edit → verify are consolidated into a single tool. Small models tend to lose coherence after three or more consecutive calls, so this cut the failure rate in half

Improvement loop**: Compilation and linting run immediately after code generation, with errors automatically fed back — eliminating the need for the model to get it right on the first try

Decomposition on failure**: If the model fails twice on the same problem, the task is broken into finer pieces, shrinking "fix a 200-line file" down to "fix just line 45"

Escalation**: Only when decomposition fails to resolve the issue is the task automatically delegated to a larger model, provided a Claude or OpenAI API key is configured. This enables a 95% local, 5% cloud operation

Token budgeting**: Rather than feeding in entire files, it summarizes and truncates so that only the critical code is presented, fitting within small models' 32k–256k contexts

Code graph**: Instead of grep-based search, it builds a symbol graph (functions, classes, call relationships), so a question like "how does authentication work?" returns only the relevant code

It's published on GitHub under the MIT license and can be installed with `npm install -g smallcode`. It supports LM Studio, Ollama, and OpenAI-compatible endpoints.

Worth watching as an effort that significantly raises the practicality of local models.

AI-Generated Slop Is Straining Bug Bounty Programs

According to the Financial Times, massive volumes of AI-generated vulnerability reports are starting to put serious strain on corporate bug bounty (reward) programs.

Unlike the era when security researchers discovered vulnerabilities by hand, reports are now mass-generated using AI tools. Companies must sift through enormous amounts of "slop" to pick out genuinely valuable reports, and their review resources are reportedly reaching their limits.

Bug bounties are, by nature, a valuable mechanism connecting external security researchers with companies — but the dynamic of AI-driven mass submissions threatening this ecosystem stands out as an unintended side effect of AI adoption.

Apple Shows Signs of Changing Its Handling of AI Coding Apps — Triggered by a Replit Update

According to Apple Insider, Apple may change how it treats AI coding apps on the App Store. The catalyst is a recent Replit update.

Until now, Apple has applied strict regulations to AI coding apps capable of running code in-app. Running code on iOS/iPadOS has historically been restricted due to security concerns. However, with Replit's latest update having passed Apple's review, deregulation of AI coding apps may now move forward.

Many in the developer community have welcomed this development, suggesting that AI-assisted development on mobile could become far more commonplace.

Vatican Establishes an AI Research Commission, Prepares to Issue Its First-Ever Encyclical on AI

According to the AP, the Pope has established a dedicated research commission on AI. At the same time, the Vatican is reportedly preparing to issue an encyclical on AI for the first time.

The Pope has repeatedly addressed the ethical use of AI and has met with technology industry leaders including Elon Musk and NVIDIA's Jensen Huang. The new commission is being positioned as a serious engagement, from the religious world, with AI's impact on society.

A Research Framework for Autonomously Planning and Executing Long-Horizon Image Editing Tasks

A study titled "From Plans to Pixels," featured on Hugging Face's Daily Papers, is attracting attention.

The research proposes an experiential learning framework that autonomously plans and executes long-horizon image editing tasks. Rather than relying on conventional static teacher imitation, it actively learns to select the optimal tools and regions through trial and error.

By combining a structured task planner with an orchestrator trained on outcome-based rewards from vision-language judges, it achieves multi-step edits with high consistency and reliability.

A fascinating study showing that agentic approaches are steadily making their way into the image editing domain as well.