Use the official ID
For the instruction checkpoint, the Hugging Face model card uses `google/gemma-4-26B-A4B-it`. A file named Q4, Q5 or Q6 is a quantized distribution of a checkpoint, not a new Google model family.
English
Gemma 4 26B A4B × SillyTavern
Gemma 4 26B in SillyTavern
The name is the first trap. Gemma 4 26B A4B is a 25.2B-parameter MoE model with about 3.8B active parameters, not a dense 26B model and not the 31B checkpoint. This guide connects the name, memory plan, template and frontend before you tune a character card.
Check the official model facts ↗
Name and hardware first
26B A4B is MoE, so the number needs context
Google lists the official checkpoint as Gemma 4 26B A4B. The model card gives 25.2B total parameters, 3.8B active parameters, 30 layers, image input and a 256K context window. Community GGUF filenames may add IT, Q4 or a publisher name, but those are distribution labels.
Plan for weights plus context
The active-parameter count helps explain speed, but it does not make every file a 4 GB model. Weight size, runtime buffers, the 550M vision encoder, KV cache and your context length all affect memory. Treat a community file size as a starting measurement.
Do not inherit the 31B recipe
31B is dense and has about 30.7B parameters. 26B A4B routes tokens through experts. A preset or hardware estimate written for Gemma 4 31B can be a poor fit for this checkpoint.
SillyTavern connection order
Connect the backend before the character card
A clean sequence prevents a template problem from looking like a bad sampler or a weak character card.
- 01
1. Confirm the exact checkpoint
Choose the provider or local file that identifies Gemma 4 26B A4B or its precise publisher variant. Copy the provider model slug instead of typing `gemma-4-26b` from memory.
- 02
2. Select the compatible API mode
For KoboldCpp, llama.cpp, LM Studio or an OpenAI-compatible provider, use the connector and endpoint documented by that backend. Start with Chat Completions when it is supported; Text Completion leaves more formatting to you.
- 03
3. Let the backend apply Gemma 4 format
Google’s format uses `system`, `user`, `model`, `<|turn>` and `<turn|>`. Use the current Gemma 4 or Gemini tokenizer and chat template supplied by the backend. Do not paste a Gemma 3 template over it.
- 04
4. Add thinking deliberately
Google documents `<|think|>` at the start of the system prompt as the thinking switch. Test it with a short prompt, then add the card and preset. Keep only the final answer in ordinary multi-turn history.
Community posts mention SillyTavern 1.17, KoboldCpp `--jinja`, the Gemma4/Gemini tokenizer and a literal `<|think|>` marker. These reports are useful troubleshooting leads, not Google defaults for every backend.
When the output breaks
Repair the layer that actually failed
Thinking markers appear as prose
Check the tokenizer, chat template and backend generation mode. If the backend already applies the template, adding raw markers can produce visible control tokens. Compare thinking on and off with one short prompt.
Thinking is inconsistent
Confirm that your SillyTavern version and loader support the Gemma 4 format. A community solution may use `--jinja` or `<|think|>`, but the correct switch depends on the server. Do not combine every workaround at once.
The card repeats or loses details
Reduce duplicated context before changing temperature. Verify the context limit and the card text, then test Google’s starting sampler values, temperature 1.0, top-p 0.95 and top-k 64, as an experiment.
A browser-side alternative
Keep the character wiki open and ask beside it
Tabbit is not a local GGUF runner or a SillyTavern card manager. It helps when the setup problem is scattered reference material: keep a character page, lore notes or a troubleshooting thread in view while you chat with a model available in the browser.
- 1
Install Tabbit
Download the Chromium-based desktop browser for macOS or Windows. This route does not require a local Gemma file, a provider key or a SillyTavern server.
- 2
Choose a model that is listed
Open the in-product model picker after installation. The current product code lists models such as GPT-5.4, Gemini-3.1-Pro, Claude-Sonnet-4.6 and DeepSeek-V4-Flash. Gemma 4 is not in that current mapping, so this page makes no promise that Tabbit runs Gemma.
- 3
Bring the source into chat
Leave the wiki or setup notes in a tab. Type `@` to reference a page, screenshot or file, then ask for a scene outline, a consistency check or a summary of the settings you just compared.
Choose the client
SillyTavern for the card, Tabbit for the source page
They solve different bottlenecks. Keep the specialized frontend when you need its controls, and use a browser route when your context lives across pages and files.
| SillyTavern | Tabbit | |
|---|---|---|
| Character cards and lorebooks | Core workflow | Reference the source page |
| Gemma 4 local checkpoint | Bring your own backend | Use a model shown in the picker |
| Template and sampler control | Detailed | Managed by selected model |
| Web context | Paste or use extensions | Reference a tab or file with @ |
| First useful reply | Endpoint + template + preset | Install + choose a listed model |
Bring the source into chat
Leave the wiki or setup notes in a tab. Type `@` to reference a page, screenshot or file, then ask for a scene outline, a consistency check or a summary of the settings you just compared.
Web context
Reference a tab or file with @
FAQ
Gemma 4 26B A4B and SillyTavern, answered
What is the official Gemma 4 26B name?+
Google’s model card calls it Gemma 4 26B A4B. The Hugging Face instruction checkpoint is `google/gemma-4-26B-A4B-it`. Community publishers may add their own suffixes.
Is 26B A4B the same as Gemma 4 31B?+
No. 26B A4B is MoE with 25.2B total and about 3.8B active parameters. Gemma 4 31B is a dense 30.7B model.
How much VRAM do I need?+
There is no single safe number. Quantized weight size, KV cache, context, runtime buffers and the vision encoder all matter. Measure the exact GGUF or backend build with your intended context instead of using the 26B label as a VRAM promise.
Why do I see `<|think|>` in the reply?+
The tokenizer or template may not match the backend, or the server may have exposed the control token as text. Check Gemma 4 formatting, tokenizer and thinking settings before changing sampling.
Can Tabbit run Gemma 4 26B?+
Do not assume it can. The current Tabbit model mapping does not include Gemma 4. Install Tabbit and use only a model shown in the current picker.
Fix the model path before the preset
For local Gemma 4 26B A4B, verify the checkpoint, memory budget, tokenizer and template in that order. For a character wiki or setup thread beside chat, install Tabbit and check the current model picker.
Available for macOS and Windows. Model availability can change.