gemma

A humorously named Gemma 4 E2B fine-tune called CRAP (Compiler Replacement for Assembling Programs) that emits working x86_64 ELF binaries directly from text prompts, skipping compilers and assemblers, now at v1.1 with a browser-based WebGPU demo.

NVIDIA released Nemotron 3.5 Bolt E4B 4GB, a post-trained Gemma 4 E4B model with multi-turn agentic and conversational capabilities optimized for running on RTX, Jetson, and DGX Spark devices, claiming ~40% higher quality and ~80% fewer output tokens than peers.

a9lim and treetowntree launched neuron lens, a browsable database and tooling for reading every neuron and attention head in Qwen3-8B and Gemma 4 12B using R-lens, S-lens, and mid-lens variants, with steering predictions correct over 90% of the time.

Google的Project Suncatcher卫星成功在太空启动TPU，机载Gemma 4模型生成了首条来自太空的回应（一首俳句）。这为2027年实现双星互联奠定了基础，最终目标是在太空中部署AI数据中心。

Google's Project Suncatcher space AI data center moonshot completes its first "hello world" test: successfully running the Gemma model on a custom-built TPU system at an altitude of about 500 miles and receiving a haiku about space in response, validating the feasibility of on-orbit AI computing.

Google's Suncatcher prototype space data center satellite is online and communicating with the ground after launching on a SpaceX Falcon 9, running Google DeepMind's Gemma 4 on onboard TPUs to perform inference in orbit. The test will assess how the chips handle space radiation and thermal stress, with two more satellites planned for 2027.

Gemma-4-31B-Humanlike-Chat 是基于 Google Gemma 4 31B 的微调模型，通过 SFT 与 on-policy 蒸馏训练，实现了无需提示的自然人类化文本风格，同时在 When2Call、BFCL 等工具调用基准上超越原版，且知识能力与基座持平。

llama.cpp has submitted a performance PR for the small Gemma 4 models E2B and E4B that makes layers reusing KV across layers compute only the last token during prefill. On an M4 Pro (Metal), this increases E2B’s prompt processing speed from about 1614 tok/s to 4603 tok/s (roughly 2.9×), while keeping the output identical.

ServiceNow AI presents Apriel Voice 27B Preview, an open speech-language model built on Gemma 4 26B-A4B that unifies speech understanding, reasoning, and tool use in a shared backbone, trained with a two-stage procedure and online distillation to reach 98.90 on Big Bench Audio and 81.13 on Speech-BFCL with 1.91s mean latency.

Speculation about upcoming Gemma 5 models, based on a tweet from Google's Oriol Sanseviero hinting at a possible release.

Google Gemma 声称利用 Gemini 4 Argon 在 A100 上通过 vLLM 实现了 2.85 倍推理加速，gemma-4-E2B-it 吞吐量从约 37,900 提升至 108,000+ tokens/sec，且保持 100% token parity、无质量下降。

This paper evaluates the 305M-parameter Gemma 4/EmbeddingGemma 2 speech encoder on Saudi and Gulf Arabic ASR without modifying its weights, comparing a frozen CTC reader and an adapter-augmented system against Cohere Transcribe Arabic and other baselines, showing competitive CER and strong efficiency gains. Code and results are publicly released.

This preprint proposes NR-SAE, a sparse dictionary method that reuses a frozen MoE router's expert selection to address per-expert autoencoder banks, showing better reconstruction than rank-addressed, candidate-matched, and randomized-identity controls on Gemma 4 26B-A4B, while noting speed and causal-faithfulness results remain untested.

A user shares their experience using the fast DiffusionGemma model locally for image analysis and offline image-generation prompting, and asks the community for other use cases where its speed tradeoff pays off.

Phil Schmid 介绍 Google 新推出的 Gemma 官方文档站点，涵盖模型能力、微调指南、推理优化以及模型工作原理的教育内容，全部集中在一个平台上。

Google DeepMind与新加坡GovTech合作，基于开源模型Gemma微调构建设备端诈骗检测系统，可在保护个人隐私的同时实时识别本地社交工程诈骗威胁。

Google DeepMind and Singapore's GovTech announced an on-device AI scam detection model, fine-tuned from the lightweight Gemma model, that flags manipulative conversations locally without compromising end-to-end encryption or user privacy. The two-stage pipeline was presented at Singapore International Cyber Week 2026 and trained in a sovereign sandbox using only anonymised scam data.

作者开源了一个类 Cotypist/Granola 的桌面应用 LokalBot，利用本地 Gemma 4 E2B（6bit 量化）基于自有会议记录实现约 50ms 延迟的打字自动补全，并在 782 个预测点上对比 Cotabby 取得略优表现。

有人将谷歌的 EmbeddingGemma 2 改造为语音识别模型，通过在冻结参数的嵌入模型上训练轻量级 Reader，把语音转写从生成问题变成检索问题，在 M3 MacBook Air 上实现 288× 实时速度（1 小时音频约 13 秒转写），WER 约 5%，可能重塑端侧 STT 市场。

Feiteng Li 分享了一份针对 EmbeddingGemma2 的开源资源清单，涵盖 DigUp、Recall 等本地多模态搜索应用、决策模型微调方案，以及 llama.cpp、MLX、LiteRT、Transformers.js 等多端部署路径。相关仓库已开源并欢迎社区贡献。