One Slider Lets a Mini PC Run 26B LLMs: Generation on the iGPU Is 21–39% Faster
Ollama kept running out of memory on a 32 GB mini PC: Windows gives the integrated Intel Arc only about half the RAM, 18 GB. The author raised the Shared GPU Memory Override limit to 26 GB, and models with 20–35 billion parameters now fit entirely. Generation on the GPU is faster than on the CPU, by 21% for Gemma 4 and 39% for Tiel-Coder, while energy per 1000 tokens drops by 28–42%. For Core Ultra owners, this can stand in for a discrete graphics card.
AI-processed from Habr AI; edited by Hamidun News
Picture this: you download the Gemma 4 26B language model, the file weighs 18 GB, your mini PC has 32 GB of RAM, and Ollama still crashes with an out-of-memory error. The reason is that Intel Arc integrated graphics has no video memory of its own. It borrows a slice of shared RAM, and by default Windows allows it roughly half.
Ollama saw only 18 GB, so models that were a tight fit would not load. The author of a Habr write-up found a single setting in Intel's software that raises this limit to 26 GB. After that, models with 20–35 billion parameters (that is what "26B" or "35B" in the names means) ran on the GEEKOM GT1 Mega mini PC, and text generation on the integrated graphics turned out to be faster than on the CPU: by 21% for Gemma 4 and by 39% for Tiel-Coder.
For anyone with a laptop or mini PC built on Core Ultra, this is a way to run large local models without buying an expensive graphics card.
One Slider Instead of a Graphics Card
The test hardware: an Intel Core Ultra 9 185H processor (16 cores, up to 5.1 GHz), Arc integrated graphics with 8 Xe cores, and 32 GB of dual-channel DDR5-5600. The "virtual video memory" limit is changed in the Intel Graphics Software app under "Settings — System — Shared GPU Memory Override." The author set it to 26 GB, more than 80% of the memory, and rebooted the computer.
For Ollama to use the integrated graphics at all, you need the variables OLLAMA_VULKAN=1 and OLLAMA_IGPU_ENABLE=1. After the reboot, Ollama reported 26.2 GB, and the ollama ps command showed "100% GPU": all of the model's layers sit on the graphics side.
The hardware did not get faster, but models of 17–18 GB now fit. Gemma with a 24,576-token context window takes 16–17 GB, and Tiel-Coder takes 17–18 GB. That leaves 5–6 GB for Windows and other programs, which was enough: OpenSCAD, Blender, SolidCraft and the arduino-cli compiler ran alongside the model.
What the Measurements Showed
The author compared four MoE models: Gemma 4 26B-A4B, Tiel-Coder 35B-A3B (a third-party fine-tune for programming, built on the Qwen 3.5 MoE architecture), GPT-OSS 20B and Qwen3-Coder 30B-A3B. An MoE model has many "experts," but only a small fraction is active for each token: in Gemma, roughly 4 billion of 25.2 billion parameters. It needs the memory of a large model but computes almost like a small one.
The tasks were engineering jobs: a gear for a Pip-Boy in OpenSCAD and FreeCAD, a handle in Blender, and an enclosure for an ESP32-C3 board in SolidCraft via MCP. For each task, the model received a reference image.
- Generation speed: Gemma 15.8 → 19.1 tokens/s (+21%), Tiel-Coder 12.2 → 16.9 (+39%).
- Average CPU package power falls from 45 to 32–33 W. Energy per 1000 tokens of output drops by 28% for Gemma and by 42% for Tiel-Coder.
- In SolidCraft, the time savings are only 5–7%: the model mostly reads a long prompt of 30–60 thousand tokens describing 13 tools. The GPU still saves energy there, about 40%.
- The device does not affect part quality: the spread between two repeats of the same task is larger than the difference between CPU and GPU.
All 26 runs together cost about 0.11 kWh, which is less than one ruble at 6.5 ₽ per kWh.
Templates Decide Everything
Without template examples (for instance, code for a 30-tooth gear given for a task about a 24-tooth gear), no model succeeded, and the author did not count this mode in the results. With a template, all four models produced all three parts correctly on the first attempt.
In that case, what decides the outcome is not the model's "intelligence" but how concisely it writes. GPT-OSS and Qwen3-Coder answer more briefly and without extra comments, so they finish twice as fast as Gemma. In SolidCraft, GPT-OSS gave the best result: 55 seconds versus 114–154 for the others. Tiel-Coder used 2.4 times more input tokens than Gemma: each new attempt sends the model the entire previous dialogue, and in SolidCraft it took 13 turns where Gemma managed with five. For simple tasks with a ready template, 8B models turned out to be even faster and more economical than the big ones: for example, the gear took Qwen3 8B 34 seconds versus 65 for Gemma 26B.
What to Do in Practice
If you have a Core Ultra with integrated Arc and 32 GB of RAM, raise Shared GPU Memory Override to 24–26 GB and enable Vulkan and iGPU in Ollama. Give the model a ready code example in the prompt, and do not resend the entire dialogue on retries.
There are caveats. This is one author's test on one machine, and the CPU-versus-GPU comparison covers only 11 runs. He measured energy with Intel RAPL counters for the CPU package only; memory, SSD, fan and power supply add another 20–30%. Absolute speed is modest, around 15–19 tokens/s, but in exchange you get a quiet, inexpensive mini PC instead of a separate graphics card.
- Nov 18, 2026Claude Code Recreates Someone Else's Video Editing Style on Your Footage for 1.1% of a Weekly Limit
- Nov 18, 2026Cloudflare open-sources Kitesurf: AI agents get by on 58 MiB of memory instead of Chromium's 399
- Nov 18, 2026A Video From One Product Photo Cost $4.25 Instead of $2.42 Because of Seedance 2.0 Defects
- Nov 18, 2026Alibaba cut Qwen4-Max's cost per task by 67%, but GPT-6.1 Sol is still cheaper: $3.27 vs. $1.27
Key Questions Answered
What are the key details on One Slider Instead of a Graphics Card?
The test hardware: an Intel Core Ultra 9 185H processor (16 cores, up to 5.1 GHz), Arc integrated graphics with 8 Xe cores, and 32 GB of dual-channel DDR5-5600. The "virtual video memory" limit is changed in the Intel Graphics Software app under "Settings — System — Shared GPU Memory Override." The author set it to 26 GB, more than 80% of the memory, and rebooted the computer. For Ollama to use the integrated graphics at all, you need the variables OLLAMA_VULKAN=1 and OLLAMA_IGPU_ENABLE=1.
What the Measurements Showed?
The author compared four MoE models: Gemma 4 26B-A4B, Tiel-Coder 35B-A3B (a third-party fine-tune for programming, built on the Qwen 3.5 MoE architecture), GPT-OSS 20B and Qwen3-Coder 30B-A3B. An MoE model has many "experts," but only a small fraction is active for each token: in Gemma, roughly 4 billion of 25.2 billion parameters. It needs the memory of a large model but computes almost like a small one.
What are the key details on Templates Decide Everything?
Without template examples (for instance, code for a 30-tooth gear given for a task about a 24-tooth gear), no model succeeded, and the author did not count this mode in the results. With a template, all four models produced all three parts correctly on the first attempt. In that case, what decides the outcome is not the model's "intelligence" but how concisely it writes. GPT-OSS and Qwen3-Coder answer more briefly and without extra comments, so they finish twice as fast as Gemma.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.