Gemma 4 E4B: from a verified GGUF file to a first chat in Pūnaha 1.0.0 AI-generated Nika narration with English captions over genuine frontend footage. Captured on 19 November 2026 in an installed, registered Windows Pūnaha 1.0.0 app. Duration: 6 minutes 28.88 seconds. Spoken narration — AI-generated Nika voice, natural style, Google AI Studio 00:00:00.000 We’ll start with the Gemma model file on Hugging Face. For this walkthrough, the GGUF file is already saved on this computer, so there’s no need to download it again. Once you have your own copy, you can bring it into Pūnaha. 00:00:25.000 Now open Pūnaha’s AI Models page and choose Add GGUF model. Select the file already on your computer, then watch the upload progress. In this recording we use a different filename because the original Gemma file was already installed. The model itself is the same verified file. On a fresh installation, you can use the filename you downloaded. 00:01:05.000 The upload finishes, and Pūnaha discovers the new file. Open the local LLM Model connection to check that a chat deployment is available and enabled. This confirms the uploaded GGUF is ready to select in a workflow. 00:01:38.000 Let’s give this deployment a clear name: Gemma E4B first chat. Then load the model into memory and refresh the status until it reads In memory. The other Gemma entries you may see were already installed; we’ll use the newly uploaded deployment for this walkthrough. 00:02:27.000 Next, open Workflows and create a blank workflow called Gemma first chat tutorial. A short description makes it easy to recognise later. We’re working in a separate practice tenant, so these steps don’t affect another tenant’s workflows. 00:03:23.000 On the canvas, connect four modules in this order: Chat input, Prompt, Language model, and Chat output. Keep the required Protection gateway enabled. Save the workflow before configuring each stage. 00:03:48.000 In the Prompt stage, the system instructions ask for short, clear answers without made-up facts. The user template passes in both the conversation history and the latest message, so follow-up questions have context. 00:04:08.000 Now select Gemma E4B first chat in the Language model stage. Set temperature to zero point two, response detail to Balanced, and the maximum response length to 256 tokens. Save the draft and run the readiness check. 00:04:40.000 Before publishing, test the saved draft with a simple question: explain what a language model does in two sentences. The run takes a little time because it is using the local model. When it completes, check that the answer is relevant and that the run passed. 00:05:35.000 The draft is ready, so publish version one inside Pūnaha. That makes the workflow available in Chat. 00:05:48.000 Open Chat and select the published Gemma first chat tutorial workflow. We’ll ask for three names for a fictional garden club. Gemma returns three suggestions, showing the published workflow can answer through the local model. 00:06:17.000 Now ask a follow-up about The Bloom Brigade. Gemma gives one playful name: Petal Pop Brigade. On-screen detail and recording notes 00:00 The browser shows the fixed Unsloth Gemma 4 E4B Instruct Q4_K_M GGUF file on Hugging Face. The page lists a 4.98 GB file and SHA-256 85a896a047553e842f25297ee5b031d64ff30147d9c4af17b1e4b394cd1fab87. That same complete hash was checked against the file already in E:\GGUF\Gemma. The file is not downloaded again in this recording. 00:25 In a separate First Chat Tutorial tenant, AI > Models shows the embedded local runtime. Add GGUF model begins a real upload and the progress indicator advances. The original filename was already installed globally, so the recording uses an alternate local filename containing the same verified model bytes: gemma-4-E4B-it-Q4_K_M-first-chat-tutorial.gguf. A normal first installation can use the original downloaded filename. 01:05 Pūnaha discovers the uploaded file. The local LLM Model connection becomes available, and the new chat deployment appears enabled. It is renamed Gemma E4B first chat for this tenant. The model is loaded; Refresh confirms In memory. The earlier already installed Gemma and EmbeddingGemma entries remain visible but are not selected for this workflow. 02:27 Workflows > Create workflow > Blank workflow creates Gemma first chat tutorial. Its description is Practice text chat with a local Gemma model. 03:23 The draft canvas is populated in order: Chat input, Prompt, Language model, Chat output. The required Protection gateway remains enabled. The workflow is saved. 03:48 The Prompt's system instructions ask for clear, brief answers and no invented facts. Its user template includes {{input.messages}} for conversation history and {{inputText}} for the latest message. 04:08 The Language model stage selects Gemma E4B first chat, which is linked to the newly uploaded GGUF. Temperature is 0.2, response detail is Balanced, and the advanced maximum is 256 response tokens. The draft is saved and the readiness check passes. 04:40 Test draft sends: In two sentences, explain what a language model does. The saved draft run completes in 21 seconds and returns: A language model is a type of artificial intelligence that is trained on vast amounts of text data. This training allows it to understand, generate, and predict human language in a coherent and contextually relevant manner. 05:35 Publish version confirms immutable workflow version 1 inside the local app. This does not publish the Docs website. 05:48 Chat selects Gemma first chat tutorial, published version 1. The first message is: Suggest three names for a fictional garden club. The reply is: The Bloom Brigade, The Verdant Vines Society, and Petal Pushers Guild. 06:17 The follow-up is: Make "The Bloom Brigade" more playful. Give one name only. The reply is: Petal Pop Brigade. The recording joins two genuine frontend captures after a recording interruption and omits the sign-in screen. It does not simulate Pūnaha screens or model answers. The published 1.0.0 package and another installation may behave differently; repeat the model load, draft run and Chat checks on your own hardware.