Edge0's New 145B AI Model Fits on a Single Gaming GPU
Startup Edge0 released a 145-billion-parameter AI model that runs on one gaming GPU instead of a data center, no cloud needed.
Most AI models that claim "frontier" performance need a rack of data-center GPUs to even load. Edge0-145B needs one graphics card you could buy at a computer store — and the company behind it says it still writes working code.
What actually shipped
On November 12, a company called Edge0 released two open-weight models, Edge0-145B and the smaller Edge0-120B, along with the software needed to run them. The headline number is 145 billion parameters — roughly the scale of models that usually demand a cluster of enterprise GPUs costing lakhs of rupees a month to rent.
The trick is a design called mixture-of-experts, or MoE. Instead of using all 145 billion parameters (the internal "knobs" a model tunes during training) for every single word it generates, the model only switches on about 10 billion of them at a time — picking whichever smaller "expert" sub-network fits the task. That's why Edge0 claims the model needs around 29GB of memory and can run its reasoning (called "inference" — the process of the model actually generating an answer) on a single well-specced consumer GPU or a capable laptop, rather than a server farm. Edge0 says decoding runs above 80 tokens per second, fast enough to feel usable for real coding work rather than a lab demo.
Worth flagging: those performance numbers come from Edge0 itself. Nobody outside the company has independently verified them yet, and early testers report Windows and Linux support needs an NVIDIA GPU with at least 16GB of video memory — Mac support is still patchy depending on which write-up you read.
The "open source" asterisk
Edge0 built the model on top of Qwen3.8-Flash-Next, an open model family from Alibaba, and is releasing the weights on Hugging Face under something called the "Edge0 Community Model License 1.0," with Qwen's own community license layered underneath. That combination has already sparked an argument online: releasing the weights (the trained numbers you can download and run) isn't the same as open source in the traditional sense, where you'd also get full rights to modify and redistribute without restriction. Developers on Hacker News have been blunt about it, calling this "open-weight," not open source — a distinction that matters a lot if you're planning to build a commercial product on top of it.
"A major turning point for our crew. Frontier-level agentic coding, now running fully on-device. Edge0-145B and Edge0-120B are available with open weights. Give them a try." — Edge0, in its release announcement
Why this matters for Indian developers
This lands differently in India than in Silicon Valley. Indian startups and solo developers routinely get priced out of frontier AI by two things: dollar-denominated API bills that swing with the rupee, and the sheer scarcity of GPU capacity in Indian data centers compared to the US. A model that runs on a single GPU you already own — or one bought through India's growing local cloud and GPU-rental scene — sidesteps both problems for coding-heavy teams.
There's a data angle too. Under India's Digital Personal Data Protection (DPDP) Act, companies handling sensitive code or customer data have to think hard about where that data travels when it's sent to a third-party AI API. Running inference entirely on a local machine means proprietary code or client data never leaves the building — a meaningful selling point for Indian fintech, health-tech, and defence-adjacent software teams who've been wary of sending code to US-hosted models.
- Size: 145B total parameters, ~10B active per token (mixture-of-experts)
- Base model: built on Alibaba's Qwen3.8-Flash-Next
- Hardware: claimed to run on one consumer GPU (16GB+ VRAM) or a high-end laptop
- Memory footprint: around 29GB peak, per Edge0's own figures
- License: Edge0 Community Model License 1.0 plus Qwen's community license — open-weight, not fully open source
None of this makes Edge0-145B a guaranteed replacement for GPT-6 or Gemini-class tools sitting in the cloud — those still win on raw capability and don't ask you to own a GPU. But the direction matters more than this one release: every few months, the GPU you need to run a genuinely capable model gets cheaper, and the gap between "what a data center can do" and "what your desktop can do" keeps shrinking. For developers anywhere compute is expensive or data has to stay local, that gap closing is the real story — Edge0 is just the latest company to prove it's happening faster than expected.
What's Your Reaction?
Like
0
Dislike
0
Love
0
Funny
0
Angry
0
Sad
0
Wow
0