Home / Companies / Google DeepMind
All Google DeepMind items under Companies on AIFIRST News, newest first. 14 items.
ai-in-browser is an open-source browser-based assistant that connects to OpenRouter, OpenAI, Claude, Gemini, and local models through a unified interface. It includes an agentic mode for multi-step tasks directly in the browser without switching tools. All provider credentials are stored locally, with no data sent to third-party servers.
ai-code-review-bot is an open-source tool that uses Google Gemini to analyze pull request diffs and post inline review comments directly on GitHub. The bot integrates with the standard PR workflow, providing automated AI code review feedback without leaving the repository.
A developer released gemma4.c, a project that implements a modern language model in 700 lines of C. The implementation runs Gemma-class models on a standard CPU without requiring GPU hardware or complex dependencies. The code is designed as an educational reference showing how modern LLM inference works at the system level.
Qwen 3.8 27B has reached 9th position on Code Arena, a benchmark ranking models by performance on real-world developer coding tasks, while Gemma 4 31B sits at 80th despite having more parameters. The gap illustrates how architecture and training choices matter more than raw parameter count for practical coding output. The result positions Qwen 3.8 27B as competitive with much larger frontier models.
A developer spent $266 and used four different AI models to gain full control of an Amazon Fire HD tablet that resists rooting. After Claude, GPT-4o, and Gemini were unable to complete the task, GLM-5.3 finished it in a single day by generating the necessary ADB commands and scripts. The post documents the process and highlights the uneven strengths of frontier models on low-level hardware tasks.
A developer fine-tuned Gemma 4 12B targeting tool calling for agentic coding use cases, achieving a 2.7x improvement over the base model. The fine-tune addresses documented weaknesses in the stock model's tool use accuracy, making it more viable for local coding assistant workflows.
nobuzz adds a /debuzz slash command to Claude Code that routes a verbose response through Google's Gemini via the Antigravity CLI with the single job of rephrasing it as plain human speech. It offers three output modes, colleague, manager, and director, each more condensed than the last, and colleague mode preserves code blocks and file paths. The repository has 139 stars and an MIT license.
RoleMux is a TypeScript npm package that lets developers define a Planner, Worker, and Reviewer agent pipeline once in a YAML manifest and run it with different AI backends (OpenAI Codex, Claude, Gemini, Hermes). It supports automatic review-and-fix loops with configurable iteration limits and runs fully local without cloud dependencies. MIT-licensed with a published npm package and GitHub Actions CI.
TrueForge is an open-source runtime framework that manages the execution loop for LLM-powered agents, handling model calls, MCP tool integration, sandboxing, approvals, context management, and session state. It supports multiple model providers including OpenAI, Anthropic, and Google Gemini, and offers a chat UI, HTTP API, and embeddable UI components. Deployment options range from local SQLite for personal use to Postgres and Redis for team deployments. The repository has 419 commits and includes benchmarking data comparing it against competing frameworks.
A hands-on technical writeup on running LLM inference across multiple AMD Radeon Pro V620 GPUs in a home server. The author found that layer parallelism hurt performance, dropping from 19-20 tokens per second on a single card to 12-13 across four cards, while speculative decoding roughly doubled speeds to about 40 tokens per second on a single card. PCIe 3.0 with 8 lanes created bottlenecks for inter-GPU communication, with peer-to-peer transfers and BIOS tuning improving tensor parallelism across two cards. Models tested include Gemma4-31B and Deepseek V4 Flash.
A 125-million-parameter decoder-only transformer was trained to autocomplete piano melodies in real time, reaching 108 note predictions per second on an iPhone 15. The key design choice was a unified note token format encoding pitch, timing, duration, and velocity together, so the model advances by complete notes rather than individual attributes. Post-training with Direct Preference Optimization using Gemini-based pairwise evaluation boosted preference scores from 24.5% to 69%, and the model powers the RollTab iOS app.
LLaMA-Factory provides a unified web UI + CLI for fine-tuning LLMs and VLMs without writing custom training code. Supports LLaMA, Mistral, Qwen, DeepSeek, Gemma, Phi, Yi, and 90+ others. Positioned as a no-code altern...
OpenUI released OUI-1, a model fine-tuned on DiffusionGemma using OpenUI-Lang, a custom domain-specific language for UI generation. The model generates bespoke interface components rather than raw HTML, targeting design-to-code workflows.
ha-llm is an open-source Home Assistant integration that exposes a single llm.ask service call routing to OpenAI, Gemini, Claude, OpenRouter, or Ollama backends. It simplifies switching LLM providers in Home Assistant automations without changing service call syntax.
AIFIRST 
