The Archive

gemma-4-31B-it-DFlash open-weights model released on Hugging Face, pending llama.cpp integration.

User documents personal journey running open-weights models locally on increasingly expensive hardware (M3 Ultra to RTX Pro 6000), testing Qwen, DeepSeek, Gemma, MiniMax.

NVIDIA releases Gemma-4-26B-NVFP4, a 4-bit quantized variant achieving 18.8GB footprint with minimal benchmark degradation vs. full precision.

Informal comparison of Qwen 3.6 27B and Gemma 4 31B on local Pac-Man gamedev task; Gemma produced faster, more logical output despite lower token throughput.

Qwen 3.6 27B/35B models outperform older ~30B alternatives (Qwen Coder, Gemma) on coding and agent tasks.

Evaluates EuroLLM, Aya Expanse, Gemma on emotion preservation in machine translation across 28 emotion categories in 5 languages.

Reddit discussion speculating on future larger MoE variants of Gemma-4 and Qwen 3.6 models.

User reports Qwen 3.6 and Gemma 4 performing expert-level work locally on consumer hardware (RTX 3090), replacing skilled human labor.

Gemma 4 chat template bug identified: JSON Schema `anyOf` patterns render as empty `type` fields, breaking tool calling across inference engines.

llama.cpp merged SM120 native NVFP4 quantization support; community released GGUFs for Gemma-4-31B and Nemotron-Cascade models.

User reports local LLMs (Qwen 27B, Gemma 4 31B) underperform Claude for coding tasks due to poor tool-calling and decision-making.

Benchmark comparison of 2026 4B-class open-weights models (Gemma 4, Qwen 3.5, Granite 4, Nemotron-3, Phi 4-mini) on M3 Pro.

Production case study: integrating small LMs (Gemma 4E2B, Qwen3 0.6B) into Android app reveals memory, latency, and quantization challenges for on-device deployment.

Tutorial: running a local coding agent with Gemma 4 and Pi using llama.cpp for on-device inference.

KL divergence benchmark comparing Gemma 4 and Qwen 3.6 with quantized KV cache compression (q8_0, q4_0).

Google announces first Austrian data center in Kronstorf, creating 100 jobs; infrastructure expansion unrelated to Gemma model development.

Comparative evaluation of Qwen3.6-27B, Qwen3.6-35B, Qwen3.5-27B, and Gemma4 on structured document task using RTX 5090.

User reports Gemma 4 e4b extracted from Android app outperforms published versions; unclear if genuine discovery or misconfiguration.

SafetyALFRED benchmark evaluates multimodal LLMs (Qwen, Gemma, Gemini) on hazard recognition and mitigation in embodied kitchen environments.

User documents that Gemma 4's default vision budget (280 tokens) is too low; Variable Image Resolution mode requires manual tuning.

Google Gemma team soliciting community input on next model variants via Twitter poll.

Interactive benchmark comparing coding capabilities: Qwen 3.6 35B, Qwen 3.5 variants, Gemma 4, GLM 4.7 Flash via racing game simulation.

Google Gemma-4-E2B's safety filters render model unusable for emergency preparedness; blocks medical, water purification, maintenance info.

User comparison of Gemma 4 26B and Qwen 3.6 35B performance on consumer GPU with anecdotal quality assessments.

Hey all, Our ML team spent some time this week getting training and deployments working for Gemma-4, and wanted to document all the things we ran into along the way. * **PEFT doesn't recognize Gemma 4's custom layers.** Google wrapped vision/audio projections in a new `ClippableLinear` class that doesn't inherit from `nn.Linear`, so PEFT refuses to attach LoRA, even for text-only fine-tuning. Fix: unwrap the wrappers after loading weights but before calling PEFT. * **SFTTrainer killed training silently.** TRL hardcodes `use_cache=False`, which breaks Gemma 4's KV-sharing attention. Loss nev...

Benchmark compares token pruning compression across Qwen3, Gemma-3, Llama-3, Aya for Korean-centric NLP with English-Korean vocabulary optimization.

Gemma 4 E2B audio transcription via MLX and mlx-vlm on macOS; practical inference recipe provided.

a quiet day lets us give due respect to the enormously successful Gemma 4 launch

A welcome update from Google!