Skip to content

Type: AI model

Gemma 4 is Google’s open-weight model family released April 2026 under Apache 2.0, built from the same research as Gemini 3, with context windows up to 256K and native vision and audio. It ships in sizes from compact effective-2B variants up to a dense 31B and a 26B mixture-of-experts. GROUNDING tracks Gemma 4 releases, benchmarks, and on-device-to-server deployment.

Recent Updates

  • 2026-06-29: Ornith-1.0 is a new open-weights model for agentic coding (Simon Willison’s Weblog) · simonwillison.netAgents Code Agents LLM Evals Open Source LLMs DeepReinforce Simon Willison Ornith-1.0 Qwen-3.5
  • 2026-06-29: Ornith-1.0: open-source agentic coding models with self-improving training (Hacker News) · github.comCode Agents Agents Tool Use LLM Evals DeepReinforce AI Gemma · Qwen · OpenHands · Claude Code · vLLM · SGLang · Transformers Harbor Terminus-2 Ornith-1.0 · Qwen-3.5 Terminal-Bench 2.1 SWE-Bench NL2Repo · OpenClaw SWE-bench Verified · SWE-bench Pro SWE-bench Multilingual SWE Atlas QnA SWE Atlas RF SWE Atlas TW ClawEval
  • 2026-06-30: LLM Confidence Reports Track Commitment More Than Correctness (cs.LG updates on arXiv.org) · arxiv.orgLLM Evals Dharshan Kumaran Gemma 3
  • 2026-07-02: Windows 11 guide to choosing a local LLM runtime and model (Все статьи подряд / Искусственный интеллект / Хабр) · habr.comOpen Source LLMs Gemma · Qwen · Qwen3.5 4B
  • 2026-07-03: BaseRT claims top-tier LLM inference throughput on Apple Silicon with native Metal (cs.CL updates on arXiv.org) · arxiv.orgApple · Qwen3 · Llama 3.2
  • 2026-07-07: Gemma 4 technical report introduces an open-weight multimodal model family (cs.CL updates on arXiv.org) · arxiv.orgOpen Source LLMs Long Context arXiv · arXivLabs · alphaXiv · CatalyzeX · DagsHub · Gotit.pub · Hugging Face · ScienceCast · CORE · Influence Flower · Gemma
  • 2026-07-07: Gemma 4 Technical Report highlights multimodal, long-context claims (alphaXiv) · twitter.comOpen Source LLMs Long Context LLM Evals Google · alphaXiv · Gemma 3 · E2B E4B
  • 2026-07-09: Agon: cross-model RL that grades reasoning by competition (cs.CL updates on arXiv.org) · arxiv.orgLLM Evals Qwen3 · Qwen3.5
  • 2026-07-13: How are linear representations learned? Exact solutions to the dynamics of abstraction (cs.LG updates on arXiv.org) · arxiv.org — DINOv3
  • 2026-07-14: EYT-Bench benchmarks multi-turn dialogue with decoupled simulation, modeling, and judging (cs.CL updates on arXiv.org) · arxiv.orgLLM Evals GPT 5.5 · DeepSeek V4 Pro
  • 2026-07-15: AINews digest: Codex usage surge, agent harness observability, and heavily quantized open models for local agents (Latent.Space) · latent.spaceCode Agents Agents Tool Use LLM Evals Open Source LLMs Long Context OpenAI · JetBrains · LangChain · PrismML · Tencent OpenMOSS Locally AI · Latent.Space Richard MacManus Addy Osmani · GPT-5.6 · Qwen 3.6-27B · Bonsai 27B Hy3 · Qwen3.5-122B-A10B · GLM-4.7-Flash · GLM-5.2 · DeepSeek-V4-Flash · Mimo v2.5 MOSS-VL-Realtime
  • 2026-07-15: Running Gemma 4 26B at 5 tokens/sec on a 13-year-old Xeon with no GPU (Hacker News) · neomindlabs.comOpen Source LLMs Agents Code Agents Google · HP · Intel Ryan Findley
  • 2026-07-16: Entity familiarity probes and refusal steering in instruction-tuned models (cs.CL updates on arXiv.org) · arxiv.orgLLM Evals arXiv · Wikipedia Grzegorz Brzezinka · Bielik PLLuM · Qwen3 · Gemma 4 12B
  • 2026-07-16: Thinking Machines Lab releases Inkling, an open-weights multimodal model (Simon Willison’s Weblog) · simonwillison.netOpen Source LLMs Thinking Machines Lab · NVIDIA · Gemma Mira Murati Simon Willison · Inkling · Inkling-Small NVIDIA Nemotron
  • 2026-07-17: Introspection Fine-Tuning improves how small LLMs report internal perturbations (cs.CL updates on arXiv.org) · arxiv.orgLLM Evals Llama 3.2 Llama-1B
  • 2026-07-20: On-Policy Delta Distillation for Reasoning Models (alphaXiv) — NAVER AI Lab Byeongho Heo Jaehui Hwang Sangdoo Yun Dongyoon Han Qwen3

FAQ

What is Gemma 4?

Gemma 4 is Google’s open-weight model family released April 2026 under Apache 2.0, built from the same research as Gemini 3, with context windows up to 256K and native vision and audio. It ships in sizes from compact effective-2B variants up to a dense 31B and a 26B mixture-of-experts. GROUNDING tracks Gemma 4 releases, benchmarks, and on-device-to-server deployment.

What does this page track?

Dated radar mentions, source links, related concepts, and builder-relevant context for Gemma 4, collected automatically by GROUNDING.

When was Gemma 4 last mentioned?

Gemma 4 was most recently mentioned in a radar update dated 2026-07-20.