Model datasheet · Gemma 4
Gemma 4 26B-A4B
Google's answer to the fast-MoE era: 26B of knowledge at 4B speed, vision, 256K window, QAT 4-bit weights, and finally an Apache license. The strongest all-rounder in the 16GB class right now.
- Vendor
- Architecture
- MoE · 25.2B total / ~3.8B active
- Context
- 262,144 tokens
- License
- Apache-2.0
- Released
- 2026-03-31
- Vision
- Yes
→ GGUF (llama.cpp / LM Studio / Ollama) → MLX (Apple Silicon)
Architecture, from its own config.json
30 layers · 25 sliding window · 5 full attention
16 attention heads over 8 KV heads (GQA 2:1), head dim 256 — its full-attention layers use a different shape again, 2 KV heads of 512, which is the shape that decides the cache at long context · mixture of experts, 3.8B of 25.2B active per token · vocabulary 262,144
Where the memory goes: weights by band above, plus a KV cache of 460MB at 8K · 940MB at 32K · 2.8GB at 128K, plus the runtime's own working memory — dominated by the output projection over that vocabulary.
Inside one layer, from its own tensors
25 × sliding-window layer
5 × full-attention layer
The green line down the right of each half is the residual: it leaves the spine above the norm, skips the operation entirely and rejoins at the +. Where attention is drawn open, q, k and v are scaled to their real output widths — a model that shares key and value heads across many query heads shows one wide box beside two narrow ones.
Around the stack: an embedding of 262,144 × 2,816, a final norm, and an output projection that reuses the embedding (tied weights).
Also in every block and not drawn, because their tensor names do not fix where they belong: layer_scalar, post_feedforward_layernorm_1, post_feedforward_layernorm_2, pre_feedforward_layernorm_2, per_expert_scale, proj. Guessing an order is the one thing this diagram will not do.
§1 Characteristics by quantization band
— Not yet rated · contributions welcome —
Q4–Q5 — 4–5 bit
14GB weights · e.g. Q4_0 , Q4_K_M , mlx-4bit + KV: 460MB at 8K · 940MB at 32K · 2.8GB at 128Kdata checked Aug 2026
- quant-degradation-community — community consensus @ Q4_0 · aggregated · 2026-08-30
Editorial derivation: QAT Q4_0 is the intended local build; small loss vs full.
data checked Aug 2026
- quant-degradation-community — community consensus @ Q4_0 · aggregated · 2026-08-30
Editorial derivation: QAT limits quant loss.
data checked Aug 2026
- quant-degradation-community — community consensus @ Q4_0 · aggregated · 2026-08-30
Editorial derivation: MoE expert weights are quant-sensitive; slight downgrade at 4-bit.
— Not yet rated · contributions welcome —
FP16 — full precision (fp16/bf16)
50GB weights · e.g. bf16 + KV: 460MB at 8K · 940MB at 32K · 2.8GB at 128Kdata checked Oct 2026
- vendor-model-card — AIME 2026 (no tools) = 88.3 @ bf16 · vendor-reported · 2026-10-04
GPQA Diamond: 82.3, MMLU-Pro: 82.6. Vendor-reported. Listed on Vectara HHEM (see the factuality cell); absent from Aider polyglot, but that leaderboard has had no new entries since 2025-10, so its absence says nothing about this model.
data checked Aug 2026
- vendor-model-card — LiveCodeBench v6 = 77.1 @ bf16 · vendor-reported · 2026-08-30
A generational leap over Gemma 3's coding — but vendor-reported only so far.
data checked Aug 2026
- vendor-model-card — Tau2 average = 68.2 @ bf16 · vendor-reported · 2026-08-30
Gemma 4 adds trained-in tool calling — fixing the family's old gap.
data checked Aug 2026
- vendor-model-card — MRCR v2 @128K = 44.1 @ bf16 · vendor-reported · 2026-08-30
The 256K window is real but retrieval quality drops well before it fills — the vendor's own number says so.
data checked Oct 2026
- eqbench-creative — EQ-Bench Creative Writing v3 rubric score (0–20) = 16.03 · aggregated · 2026-10-04
Independent; row google/gemma-4-26B-A4B-it. Rubric rank 65 of 140 (board median 15.96); by Elo (1304.6, relative, judged pairwise by Claude Sonnet 4.6) it ranks 94. The rubric is the stabler figure and is what is stored. Quant and mode not stated (presumed full-precision API). Data file behind https://eqbench.com/creative_writing.html.
- eqbench-creative — EQ-Bench Creative Writing Longform score (0–100) = 50.7 · aggregated · 2026-10-04
Rank 92 of 140 on the longform board (median 59.1). Mid-table prose: solid rubric scores, weaker in pairwise comparison and at chapter length.
data checked Oct 2026
- vectara-hallucination — HHEM hallucination rate % (lower is better) = 5.2 @ bf16 · aggregated · 2026-10-04
Independent; row google/gemma-4-26b-a4b-it, README 'Last updated on September 22, 2026' (rows present since at least the 2026-05-11 README commit). Factual consistency 94.8%, answer rate 99.8%. Summarisation faithfulness at temperature 0 via API; mode not stated. One of the lower rates on the board, level with Qwen3-8B (4.8) and Qwen3-14B (5.4). MMLU-Pro 82.6 on Google's card supports the knowledge side.
§2 Known issues & what fixes them
The 256K window oversells it: retrieval accuracy at 128K is only 44.1% (vendor's own MRCR v2 number). Don't stuff the window and expect recall.
Retrieve narrowly instead of filling the context — smaller, relevant chunks beat a full window.
Evidence · 1 source · anecdotal
- Gemma 4 model card (MRCR v2 128K: 44.1) (vendor, 2026-08-30)
Early user reports of the Gemma-family habit: hallucinated names and specific details on open-ended general-knowledge questions. This is a different task from the one Vectara HHEM measures — there, summarising a supplied document, it hallucinates in 5.2% of summaries, one of the better rates on the board. Grounded in a document it is reliable; asked to recall obscure specifics from memory, less so.
Ground factual work until independent measurements arrive.
Evidence · 2 sources · anecdotal
- Gemma 4 factual hallucination reports (other, 2026-08-30)
- Vectara HHEM leaderboard (google/gemma-4-26b-a4b-it: 5.2% hallucination rate, updated 2026-09-22) (leaderboard, 2026-10-04)
An open llama.cpp report (2026-09-29, Gemma 4 26B MoE, Vulkan) describes unstable streaming tool calls when arguments span several lines or the model uses unquoted keys — edits and file writes come out malformed. The parser maintainer is reworking the Gemma 4 parser (PR #29307, open) but was sceptical of the reporter's patch, and the reporter could not reproduce the parse failures on master on 2026-10-02. Treat as unconfirmed, and as a llama.cpp parser issue rather than a model defect.
Track llama.cpp for the Gemma 4 parser rework (PR #29307). Until then, prefer non-streaming tool calls in agent harnesses if multi-line arguments misparse — a workaround inferred from the report, not maintainer-verified.