GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
GotoAI-Inc
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
GotoAI-Inc
Model Tree
google/gemma-4-26B-A4B-it
Input Modalities
Output Modalities
Supported Functionality
Weights were quantized from bfloat16 to int4, group size 64, symmetric, weight-only
(activations stay 16-bit) using llmcompressor.model_free_ptq. No calibration data was
used and the model was never loaded — the quantizer operates directly on the safetensors.
Architecture, tokenizer, chat template and processor config are the vendor's, unmodified.
11,725 Linear modules were converted, covering 83.1% of the output's bytes:
| component | precision | source | quantized |
|---|---|---|---|
| MoE experts (128 per layer × 30 layers) | int4 g64 | 45.68 GB | 12.13 GB |
| attention projections (115 modules) | int4 g64 | 2.22 GB | 0.59 GB |
| shared-expert MLP (90 modules) | int4 g64 | 1.07 GB | 0.28 GB |
embed_tokens (tied to the output head) | bfloat16 | 1.48 GB | 1.48 GB |
vision tower + embed_vision | bfloat16 | 1.15 GB | 1.15 GB |
| routers | bfloat16 | 0.02 GB | 0.02 GB |
| norms, layer scalars | bfloat16 | 0.001 GB | 0.001 GB |
| total | 51.61 GB | 15.65 GB |
This is the largest reduction in the collection for a simple reason: 88.5% of the source checkpoint is expert weight, and essentially all of it quantizes.
Group size 64, not the usual 128. The expert down_proj takes a 704-wide input and the
shared MLP's takes 2112; neither is divisible by 128. At the default group size those
layers cannot be quantized and the build collapses to a few percent of bytes. 704 and 2112
are both divisible by 64.
Left at bfloat16:
vision_tower, embed_vision — the tower's intermediate_size is 4304, not
divisible by 64, so int4 Marlin-style kernels cannot serve it; vLLM's Gemma 4 loader has
an explicit guard for this case.router.proj is built in vLLM as a GateLinear that takes no
quant_config at all and emits fp32 logits, because the top-k kernel needs fp32 for
stable routing. A quantized router would simply fail to load. It is 0.02 GB across all
30 layers, so there is nothing to gain either.embed_tokens — precision-sensitive, and it is the output head here
(tie_word_embeddings: true; no lm_head tensor exists).This model has no audio tower (audio_config: null) — Gemma 4 ships audio only on
E2B, E4B and 12B.
The source stores experts fused as 3-D tensors (experts.gate_up_proj (128, 1408, 2816),
experts.down_proj (128, 2816, 704)). model_free_ptq splits them into per-expert 2-D
weights before quantizing, so this checkpoint ships 11,520 individually quantized expert
modules (…experts.{id}.gate_proj.weight_packed, up_proj, down_proj) rather than
fused 3-D blocks. vLLM handles both layouts explicitly — its Gemma 4 loader carries a
dedicated path for "CompressedTensors-format AWQ/W4A16 _packed, _scale" expert names —
so no conversion is needed. It does mean the tensor count is high (35,923) and the index
file is correspondingly large.
bash
vllm serve GotoAI-Inc/gemma-4-26B-A4B-it-W4A16 \--max-model-len 131072 \--enable-auto-tool-choice --tool-call-parser gemma4 \--reasoning-parser gemma4
Do not pass --quantization; compressed-tensors is detected from config.json. The int4
W4A16 scheme uses Marlin kernels and runs on compute capability 7.5 and above.
Version note. Gemma 4 needs transformers >= 5.10.1, but vLLM ≤ 0.27.1 cannot serve
Gemma 4 with transformers >= 5.15: at 5.15 the config turns global_head_dim into a
per-layer head_dim override, and older vLLM reads head_dim globally, raising
AmbiguousGlobalPerLayerAttributeError at engine-config time. Use either
transformers < 5.15 with vLLM 0.25.1–0.27, or vLLM >= 0.28, which handles both
layouts.
30 layers: 25 sliding-attention (window 1024, 8 KV heads, head_dim 256) and 5 global
(every 6th, 2 KV heads, global_head_dim 512). The sliding layers are bounded by the
window at ~0.2 GB per sequence no matter how long the context; only the 5 global layers
grow, at ~20 KB/token — unusually cheap:
| context | KV cache | + weights |
|---|---|---|
| 32k | ~0.9 GB | ~16.5 GB |
| 128k | ~2.8 GB | ~18.5 GB |
| 256k (max) | ~5.4 GB | ~21.1 GB |
That is what makes a 24 GB card viable at full context. --language-model-only frees the
1.15 GB tower (vLLM skips tower weights when every modality limit is zero) plus the
multimodal profiling headroom if you need more. Note that only ~3.8B of the 25.2B
parameters are active per token, so throughput is far better than the footprint suggests.
This is arithmetic from config.json, not a measured deployment.
Speculative decoding has a vendor drafter, google/gemma-4-26B-A4B-it-assistant
(Gemma4AssistantForCausalLM, 4 layers). vLLM normalizes it to its gemma4_mtp path,
which produces one draft token per forward:
bash
--speculative-config '{"method": "mtp", "model": "google/gemma-4-26B-A4B-it-assistant", "num_speculative_tokens": 1}'
Not smoke-tested here.
The chat template defaults enable_thinking to false, so this model does not think
unless asked. Both knobs are template variables passed through chat_template_kwargs:
json
{"chat_template_kwargs": {"enable_thinking": true}} // injects <|think|> into the system turn{"chat_template_kwargs": {"preserve_thinking": true}} // keep thinking on tool-call turns
With thinking on, reasoning is emitted as <|channel>thought … <channel|>, which
--reasoning-parser gemma4 splits into reasoning_content. Per the base model card,
thinking from earlier turns should not be replayed into history — except on tool-call
turns, which is what preserve_thinking keeps.
Sampling, per the base model card: temperature=1.0, top_p=0.95, top_k=64. Place image
content before the text in a prompt. The visual token budget is configurable
(70/140/280/560/1120, default 280) — lower it for video and captioning, raise it for OCR
and document parsing.
Built with llm-quantizer:
bash
./llmq.py run --profile gemma-4-26b-a4b-it
which re-shards the source — it ships as 2 shards, the larger 49.9 GB — into 15 pieces of ~4 GB, then:
python
# llmcompressor==0.13.1a20260814, compressed-tensors==0.18.1a20260818,# transformers==5.15.1, torch==2.13.0from llmcompressor import model_free_ptqmodel_free_ptq(model_stub="gemma-4-26B-A4B-it-resharded",save_directory="gemma-4-26B-A4B-it-W4A16",scheme="W4A16",group_size=64,ignore=["re:.*vision.*", "re:.*audio.*", "re:.*router.*","re:.*layernorm_\\d+$", "lm_head", "re:.*embed_tokens.*"],device="cuda:0",)
A job holds one shard at a time, so after re-sharding the build peaks at a few GB of VRAM rather than the model size — no large GPU required.
The re:.*layernorm_\d+$ entry is not optional. compressed-tensors auto-skips norms with a
literal module_name.endswith("norm") test, which Gemma 4 MoE's suffixed norms miss —
post_feedforward_layernorm_1, post_feedforward_layernorm_2 and
pre_feedforward_layernorm_2, 90 one-dimensional tensors in all. Without that pattern they
reach the quantizer and it aborts with expected 2D linear weight.
No benchmarks have been run. Data-free round-to-nearest quantization degrades quality more than a calibrated (GPTQ/AWQ) or QAT build; how much, for your task, is unmeasured here. Treat published Gemma 4 26B A4B numbers as describing the bf16 model, not this one.
Two reasons to be more careful than usual with an MoE at 4 bits: the routers stay 16-bit here, so expert selection is unchanged, but every expert's weights are quantized without calibration, and rarely-activated experts get no more attention than hot ones. If you measure anything, measure it on your own traffic.
Apache 2.0, inherited from the base model — see LICENSE and Google's
Gemma 4 license page. The base
repository ships no LICENSE file, so the Apache-2.0 text is included here for
redistribution. "Gemma" is Google's mark; this repository is not endorsed by or
affiliated with Google.