Running Gemma-4 26B at 124 tokens/SEC on a CPU, no GPU
I wanted to know how fast a 26B mixture-of-experts model could run on a desktop CPU with no GPU. Got ~40 tok/s single-stream (lossless) and ~124 batched. The surprising part was the byte budget: for this model you compress the output head (32% of per-token bytes), not the experts (16%). The writeup has the bandwidth roofline and the dead-ends; the repo has the reproducible recipe. Happy to answer questions. Repo: https://github.com/arun-prasath2005/gemma4-cpu-moe
Running Gemma-4 26B at 124 tokens/SEC on a CPU, no GPU launched on June 30, 2026 with 10 votes, #1,316 of 22,647 launches that month and more than 90% of that year's launches.
In plain words
This project demonstrates running a 26-billion-parameter mixture-of-experts language model on a desktop CPU without a GPU, achieving approximately 40 tokens per second in single-stream mode and 124 tokens per second in batched operation. Intended for developers interested in CPU-based inference optimization, it documents bandwidth constraints and compression strategies, revealing that compressing the output head rather than the experts provides the most efficiency gains. The repository includes reproducible implementation details and technical analysis of performance bottlenecks.
written from the facts on this page · September 2026
From the sources
A 26B mixture-of-experts model on a desktop CPU with no graphics card: ~40 tokens/sec single-stream lossless, ~124 batched. The recipe, the byte budget, and where the walls actually are.
A 26-billion-parameter model on a plain desktop with no graphics card. The recipe, and the map of where the walls actually are. I wanted to see how fast a 26B mixture-of-experts model runs on a normal desktop with no graphics card. Just the CPU: an i9-13900K, 64GB of plain DDR5. I went looking because of a question that keeps deciding the architecture of the robot brain I'm building, how fast can a model think on hardware you don't rent. The answer turned into its own project. About 40 tokens/sec single-stream, lossless, or about 124 if you batch a few requests. For a 26B model with no GPU, that's a little wild. One thing up front, because it surprised me. The instinct with a mixture of…from apeg.dev
Does a similar job
all alternatives →
Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series MacJul 2026 · github.com · ▲919Hi HN, I built a specialized inference engine for running 4-bit Gemma 4 26B-A4B-IT on any M-series Mac using about 2 GB of RAM. It is called TurboFieldfare and is written in Swift and Metal. I have always adored on-device AI. It feels like magic that you can run a powerful NN on your Mac or iPhone. So I wanted to push the limits a bit and run a model whose weights don’t fit in memory. The model’s 4-bit quantized weights occupy roughly 14 GB, which makes running it with conventional inference tools almost impossible on an 8 GB or even 16 GB Mac once the OS, applications, and KV cache are…
Speeding up LLM inference 2x times (possibly)2024 · asciinema.org · ▲419Here's a project I've been working on for the last few months. It's a new (I think) algorithm, that allows to adjust smoothly - and in real time - how many calculations you'd like to do during inference of an LLM model. It seems that it's possible to do just 20-25% of weight multiplications instead of all of them, and still get good inference results. I implemented it to run on M1/M2/M3 GPU. The mmul approximation itself can be pushed to run 2x fast before the quality of output collapses. The inference speed is just a bit faster than Llama.cpp's, because the rest of implementation…
Cactus Hybrid: We taught Gemma 4 to know when it's wrongJul 2026 · github.com · ▲191Hey HN, Henry & Roman here from Cactus. A small, on-device model is fast and private, but sometimes wrong, but frontier models are getting expensive pretty fast. So, we post-trained Gemma 4 E2B post-trained to know when it's wrong. Every response comes with a confidence score between 0 and 1. Developers can accept the on-device when it's high, hand off to a bigger cloud model when it's low. By routing only 15-35% of queries to Gemini 3.1 Flash-Lite, Gemma-4-E2B matches Gemini 3.1 Flash-Lite on most benchmarks. - ChartQA: 15-20% - LibriSpeech: 25-30% - MMBench, GigaSpeech, MMAU: 30-35% -…
Gemma 4 Multimodal Fine-Tuner for Apple SiliconApr 2026 · github.com · ▲235About six months ago, I started working on a project to fine-tune Whisper locally on my M2 Ultra Mac Studio with a limited compute budget. I got into it. The problem I had at the time was I had 15,000 hours of audio data in Google Cloud Storage, and there was no way I could fit all the audio onto my local machine, so I built a system to stream data from my GCS to my machine during training. Gemma 3n came out, so I added that. Kinda went nuts, tbh. Then I put it on the shelf. When Gemma 4 came out a few days ago, I dusted it off, cleaned it up, broke out the Gemma part from the Whisper…
- ATA tiny LLM running at 21,000 tok/s on a $250 FPGA (Live Demo)Aug 2026 · mikeayles.com · ▲79
A 3.16M-parameter INT4 transformer running entirely in the on-chip memory of a Xilinx Kria KV260. Zero DRAM in the token loop, 59,965 tok/s on the fabric, bit-exact. Chat with it live.
- TCTensorDock Core GPU Cloud – GPU servers from $0.29/hr2022 · tensordock.com · ▲147
Hello HN! I’m Jonathan from TensorDock. After 7 months in beta, we’re finally launching Core Cloud, our platform to deploy GPU virtual machines in as little as 45 seconds! https://www.tensordock.com/product-core Why? Training machine learning workloads at large clouds can be extremely expensive. This left us wondering, “how did cloud ever become more expensive than on-prem?” I’ve seen too many ML startups buy their own hardware. Cheaper dedicated servers with NVIDIA GPUs are not too hard to find, but they lack the functionality and scalability of the big clouds. We thought to…
More dev tools this month
the category →
Contribute to oosexgang/hackios-xnu-kernel development by creating an account on GitHub.
Dev tools · 1d ago · github.com
open-source IDE for thoughtful software design. Contribute to devdotfast/whiteboard development by creating an account on GitHub.
Dev tools · 29d ago · github.com
Contribute to reladraw/reladraw development by creating an account on GitHub.
Dev tools · 27d ago · github.com
OpenBot▲362Grok Bot alternative: free, local, open-source, multiplayer
Dev tools · 24d ago · openbot.run
Talanoa▲346Hi HN, I’m a fullstack developer, and honestly, I’ve lost count of how many side projects I’ve worked on over the years. But one idea kept coming back to me: a different kind of email client. It first came to me in 2014: “What if we could handle emails more like conversations?” Every year, I’d revisit this concept in my “Build Notes.” Ten years later, no one had built exactly what I had in mind — so I finally decided to do it myself. That’s how Talanoa was born. I wanted something focused on people, not timelines. I was tired of marking messages as unread just to remember to reply later, or…
Dev tools · 22d ago · talanoa.email
Source: Show HN post ↗
Launched alongside, June 2026
the whole month →
Fundraisly▲1,537AI fundraising agent that finds investors and books meetings
AI · Jun 2026 · fundraisly.com
Homebrew 6.0.0▲1,481Today, I’m proud to announce Homebrew 6.0.0. The most significant changes since 5.1.0 are a new tap trust security mechanism, the new faster, smaller, default internal Homebrew JSON API, sandboxing on Linux, better defaults informed by our user survey, many brew bundle improvements, improved performance and initial support for macOS 27 (Golden Gate). Happy to discuss any questions here!
Dev tools · Jun 2026 · brew.sh
Hi HN! I was super frustrated at apps like tiktok and wechat preventing me from saving or downloading the animated gifs, so I made a little tool to turn a screen recording into a slack-sized emoji. Hope you get some value from this!
Life & fun · Jun 2026 · vorpus.github.io
Search 18 years of Hacker News and chart how any topic, tool, or person trended. Overlay terms to compare their rise and fall across 45M posts and comments.
Life & fun · Jun 2026 · hackernewstrends.com