Gemma4 怎么选:E4B vs 26B 核显实测对比

Gemma 4 家族里,E4B(7.5B)和 26B(25.2B,激活 4B)是核显用户最纠结的两款:都是 MoE、都在核显上实测跑通、生成速度还几乎一样。那是不是闭眼买大的就行?实测数据告诉我没那么简单。

同机对比:Intel Arc 140T 核显 / 23.5GB 内存 / llama.cpp b10621 / llama-bench pp512·tg128 / 8 线程。26B 为调显存后 Vulkan 全量 offload 成绩。2026-09 实测。

硬碰硬

项目gemma-4-E4Bsupergemma4-26B
参数量7.52 B25.23 B(26B.A4B)
文件体积(Q4_K_M)4.62 GiB15.63 GiB
能否整体进核显✅ 天生能⚠️ 需调显存后
生成 tg12820.34 t/s21.38 t/s
预填充 pp512380.8 t/s243.6 t/s
内存门槛8G 机器可跑32G 机器才舒服

柱状图对比生成速度:

生成速度 tg128(t/s)
26B(调显存后)21.4
E4B20.3

三个关键发现

1. 生成速度几乎打平(20.3 vs 21.4 t/s)

两款都是 MoE,解码瓶颈都在内存带宽——激活参数量级相近(E4B 激活约 1-2B,26B 激活 4B),速度自然拉不开。为 21.4 vs 20.3 的 5% 差距多背 11GB 文件,不划算。速度从来不是选 26B 的理由。

2. 但预填充 26B 反而更慢(244 vs 381 t/s)

预填充要遍历全部参数权重,26B 权重文件是 E4B 的 3.4 倍,即使激活只差几倍,读取成本也把它拖下来。Agent 场景(长系统提示词反复 prefill)反而是 E4B 的主场

3. 内存门槛天壤之别

E4B 4.6GB 文件 + 运行开销,8G 内存的小主机就能跑;26B 15.6GB + KV + 临时区,32G 才舒服,16G 机器连文件都塞不下。选 26B 之前先看自己的内存预算。

质量差距有多大

这没法用跑分量化,只能看产出。我在8 模型小说横评里测过:26B 档(supergemma4-26b)文笔金句密度明显高于 E4B 档,E4B 风格更「规整」但记忆点少。翻译、总结这类任务差距缩小。

怎么选

你的情况推荐
8~16G 内存 / 小主机 / NASE4B
Agent / 长系统提示词 / RAGE4B(prefill 更快)
32G 内存 + 愿意调显存26B(质量优先)
创作、长文写作26B
快速试水 / 不折腾E4B(下载小、门槛低)

一句话:E4B 是日用甜点,26B 是质量上限。内存够、愿意为质量多花 11GB,就上 26B;否则 E4B 的速度、体积和门槛让它更值得常驻。

相关阅读:26B 要整体进核显,需要先调共享显存注册表(教程);想让它当服务常驻,看llama-server 服务化

In the Gemma 4 family, E4B (7.5B) and 26B (25.2B, 4B active) are the two that iGPU users agonize over: both MoE, both verified running on an iGPU, decode speeds nearly identical. So just get the bigger one? The measurements say it's not that simple.

Same machine: Intel Arc 140T / 23.5GB RAM / llama.cpp b10621 / llama-bench pp512·tg128 / 8 threads. 26B uses post-tuning Vulkan full offload. Tested 2026-09.

Head to head

Itemgemma-4-E4Bsupergemma4-26B
Parameters7.52 B25.23 B (26B.A4B)
File size (Q4_K_M)4.62 GiB15.63 GiB
Fits iGPU entirely?✅ natively⚠️ after VRAM tuning
Decode tg12820.34 t/s21.38 t/s
Prefill pp512380.8 t/s243.6 t/s
RAM barruns on 8GBcomfortable on 32GB

Decode, as bars:

Decode tg128 (t/s)
26B(调显存后)21.4
E4B20.3

Three key findings

1. Decode is a tie (20.3 vs 21.4 t/s)

Both MoE — decode is bandwidth-bound, and active parameter counts are comparable (E4B ~1-2B, 26B 4B), so speed doesn't stretch. Paying an extra 11GB for 21.4 vs 20.3 (a 5% edge) is a bad deal. Speed was never a reason to pick the 26B.

2. Yet the 26B prefills slower (244 vs 381 t/s)

Prefill walks all weights; the 26B's file is 3.4× E4B's, and even with comparable active params the read cost drags it down. Agent workloads (long prompts, constant prefill) are actually E4B's home turf.

3. The RAM threshold is a canyon

E4B's 4.6GB file plus overhead runs on an 8GB mini host; the 26B needs 32GB to be comfortable — 16GB machines can't even hold the file. Check your RAM before reaching for the 26B.

How big is the quality gap

This can't be scored — only read. In my 8-model fiction benchmark, the 26B tier (supergemma4-26b) clearly outwrites E4B-tier models; E4B is more 'orderly' but less memorable. For translation and summarization the gap narrows.

Which to pick

Your situationPick
8–16GB RAM / mini host / NASE4B
Agent / long system prompts / RAGE4B (faster prefill)
32GB RAM + willing to tune VRAM26B (quality first)
Creative / long-form writing26B
Quick try-out / no fussE4B (small download, low bar)

In one line: E4B is the daily driver, 26B is the quality ceiling. If your RAM is ample and 11GB is fine for better writing, go 26B; otherwise E4B's speed, size and low barrier make it the resident pick.

Related:To fit the 26B in the iGPU, tune the shared-memory registry first (guide); to keep it resident as an API, see llama-server serving.
返回文章列表