Gemma 4 E2B

Gemma 4 E2B在没有工作指引时会幻想文件内容并声称完成任务，但加入不到80字的工作指引后能正确完成修复Bug的任务

Gemma 4 E2B是Google于2025年发布的开源轻量级模型，仅有20亿参数，可在消费级硬件上本地运行

Gemma 4 E2B comprehensively surpasses Gemma 3 27B on reasoning, coding and agentic tasks; E4B consistently beats Gemma 3 27B on every benchmark.

On the LiveCodeBench benchmark, Gemma 4 E2B scores 44, while Gemma 3 27B scores 29.

E2B offers two quantization formats: Q4O (general-purpose) and mobile quant (aggressive, including static activation quantization, per-channel quantization, targeted 2-bit quantization, and embedding/KV cache quantization).

The problem of DPO training previously crashing for Gemma 4 E2B and E4B (including Nemotron versions) has been fixed.

In the delayed package case, the model's draft leaked internal phrasing such as "in the past we have offered...", proving that human review of every draft is necessary.

Little Kiln Support Copilot is a lightweight retrieval-augmented generation (RAG) system based on local open-source models, written in Python, running two open-weight models, EmbeddingGemma and Gemma 4 E2B, locally via Ollama.

Gemma 4 E2B is a model built with quantization-aware training (QAT) (e2b-it-qat), about 4GB in size, and can run on an ordinary laptop without a GPU.

After lowering temperature to 0.2 and instructing in the prompt to apply the solution from the most similar case, the model's draft for "the vase arrived broken and I want a refund" stably adopted the damaged-item policy (free replacement within 2 days, no need to return the damaged item, provide a photo of the damage) instead of incorrectly applying the no-reason returns policy.

When the retrieval threshold was set to 0.5, a new question about a teapot pulled in irrelevant cases, causing Gemma to claim the shop did not have teapots (fabricated information); after raising the threshold to 0.6, the model states that the shop owner needs to confirm when there are no relevant cases.

Laya-Heart is a locally running dating-chat assistant tool: after the user pastes the other person's message, it displays probability-judgment cards (1-3 follow-up questions or 1-10 danger levels) and generates candidate replies, without connecting to chat software or sending anything automatically.

Gemma 4 E2B was LoRA-fine-tuned and exported as Q4_K_M GGUF (3.4 GB), run on CPU by llama.cpp's llama-server, to draft two candidate replies and filter memories.

After changing the header identification to have Gemma 4 E2B fill only the three required gaps of date, quantity, and product, all 12 header sets were correct with 0 errors, but the author cautions that the same batch of data was also used for debugging, so 12/12 should be regarded as an optimistic estimate.

On the same Google Cloud single-card T4, Gemma 4 E2B decodes at 115.3 tokens/s when cold at 45°C and drops to 106.8 tokens/s when hot at 77°C, a sustained speed drop of about 7.4%.

The experiment uses an n1-standard-2 VM, a single passively cooled Tesla T4, driver 615.71.09, vLLM 0.29.0, with the model being Gemma 4 E2B at 4-bit weights, single request, greedy decoding, 256 tokens per output.

In the speculative decoding A/B comparison, cooling the card to 50°C before each server start kept the speedup stable at 1.297x or 1.298x regardless of run order; without pre-cooling, the speedup falls between 1.279x and 1.337x.

Anvil's chat feature uses the Gemma 4 E2B model, executed through the LiteRT LM runtime.

The E2B in Gemma 4 E2B stands for Effective 2 Billion, meaning the effective parameter count is about 2 billion, and the actual parameter count may be slightly higher due to architectural design.

On-device small models still have a clear gap with cloud-based large models on complex reasoning tasks; Gemma 4 E2B is suitable for lightweight conversation and instruction following, but its capability boundary is obvious in multi-step logical reasoning or long-context understanding.