Gemma 4 E2B
An on-device open model released by Google, optimized for the TPU in the Tensor G5 chip of the Pixel 10 series. With approximately 2 billion parameters, it supports multimodal inputs including text, images, and audio, and can run entirely offline.
Timeline (last 90 days)
Anvil's chat feature uses the Gemma 4 E2B model, executed through the LiteRT LM runtime.
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.
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 the GSM8K mathematical reasoning test, Qwen3-4B scores 0.915 and Gemma 4 E2B scores 0.911, a difference of +0.4 percentage points with a confidence interval crossing zero, making it the only model on par with Gemma.
Gemma 4 E2B occupies 8.97 GiB of memory on this TPU, making it one of the candidate models with the largest weights.
On the BFCL tool-calling test, SmolLM3-3B scores 0.945, slightly above Gemma 4 E2B's 0.927, but the difference is not significant; Qwen3-4B (0.940), Qwen3-1.7B (0.935), Qwen2.5-1.5B (0.910), and Granite 3.3 2B (0.920) are all at the same level as Gemma.
In the 16-concurrency, 256-token-output test, Gemma 4 E2B achieves an output speed of 1887 tokens/second, and the models faster than it include Qwen2.5-0.5B (4176, 2.21x), Qwen3-0.6B (3210, 1.70x), Qwen2.5-1.5B (2780, 1.47x), and Qwen3-1.7B (2049, 1.09x).
In the scenario where 16 requests each carry a 3,072-token long prompt, Gemma 4 E2B maintains 1049 tokens/second, while Granite 3.3 2B drops to 732 (0.70x), SmolLM3-3B drops to 725 (0.69x), Phi-4-mini drops to 404 (0.39x), and Qwen3-4B reaches only 373 (0.36x).
Gemma 4 E2B's KV cache capacity is 321,280 tokens, which is 2.6 to 6.3 times that of Granite 3.3 2B (125,184), SmolLM3-3B (120,320), Qwen3-1.7B (100,096), and Qwen3-4B (50,688).
Gemma 4 E2B stores only one KV head per layer, and 28 of its 35 layers retain only a 512-token sliding window, with 20 layers reusing the cache of earlier layers.
7 more timeline events
All Facts (20)
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
85%UnverifiedGemma 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
75%UnverifiedOn-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.
50%UnverifiedAnvil's chat feature uses the Gemma 4 E2B model, executed through the LiteRT LM runtime.
50%UnverifiedThe 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.
50%UnverifiedGemma 4 E2B's KV cache capacity is 321,280 tokens, which is 2.6 to 6.3 times that of Granite 3.3 2B (125,184), SmolLM3-3B (120,320), Qwen3-1.7B (100,096), and Qwen3-4B (50,688).
50%UnverifiedOn the BFCL tool-calling test, SmolLM3-3B scores 0.945, slightly above Gemma 4 E2B's 0.927, but the difference is not significant; Qwen3-4B (0.940), Qwen3-1.7B (0.935), Qwen2.5-1.5B (0.910), and Granite 3.3 2B (0.920) are all at the same level as Gemma.
50%UnverifiedIn the 16-concurrency, 256-token-output test, Gemma 4 E2B achieves an output speed of 1887 tokens/second, and the models faster than it include Qwen2.5-0.5B (4176, 2.21x), Qwen3-0.6B (3210, 1.70x), Qwen2.5-1.5B (2780, 1.47x), and Qwen3-1.7B (2049, 1.09x).
50%UnverifiedIn the scenario where 16 requests each carry a 3,072-token long prompt, Gemma 4 E2B maintains 1049 tokens/second, while Granite 3.3 2B drops to 732 (0.70x), SmolLM3-3B drops to 725 (0.69x), Phi-4-mini drops to 404 (0.39x), and Qwen3-4B reaches only 373 (0.36x).
50%UnverifiedGemma 4 E2B stores only one KV head per layer, and 28 of its 35 layers retain only a 512-token sliding window, with 20 layers reusing the cache of earlier layers.
50%UnverifiedGemma 4 E2B occupies 8.97 GiB of memory on this TPU, making it one of the candidate models with the largest weights.
50%UnverifiedOn the GSM8K mathematical reasoning test, Qwen3-4B scores 0.915 and Gemma 4 E2B scores 0.911, a difference of +0.4 percentage points with a confidence interval crossing zero, making it the only model on par with Gemma.
50%UnverifiedThe same Gemma 4 E2B model varies from 34.9 to 131.9 tokens/s in single-request generation speed across different AWS inference backends, a significant gap.
50%UnverifiedThe int4 build from Google's QAT checkpoint on Gemma 4 E2B is less than 0.5 points lower in accuracy than bf16 (0.678 vs 0.683), while the int8 build rounded directly from bf16 dropped 1.3 points.
50%UnverifiedAfter quantizing Gemma 4 E2B's embedding table to int4, the model size dropped from 6.32 GiB to 2.85 GiB, with the reduction coming mainly from the embedding table rather than the Transformer body.
50%UnverifiedGemma 4 E2B's embedding table occupies the vast majority of the model weights, because the vocabulary usually has tens of thousands to hundreds of thousands of entries, each with a dimensionality of several thousand, causing the embedding table to easily occupy several GiB, so the quantization strategy must prioritize the embedding.
50%UnverifiedIn terms of inference speed, Gemma 4 E4B takes approximately 0.9 seconds per photo and E2B approximately 0.6 seconds, completing 165 photos within three minutes.
50%UnverifiedCRN v2 is placed on top of the frozen Gemma 4 E2B model, with the base weights unchanged throughout.
50%UnverifiedA model with 2B parameters typically requires only about 1–2 GB of memory after INT4 quantization, enabling it to run on mainstream smartphones, embedded development boards, and similar devices.
50%UnverifiedGemma 4采用混合专家架构（MoE），E2B后缀表示Effective 2B，每次推理只激活约20亿有效参数
50%