A 145-Billion-Parameter Model Just Moved Onto a Laptop, at 29 Gigabytes and 80 Tokens a Second, and It Is Open Source
The tradeoff local AI was supposed to force
On-device AI has always asked for a compromise. Run a small model on your own hardware and you keep your data, you avoid a per-token bill, and you can work offline, but you give up capability on hard coding, reasoning, and agent work. Run a frontier model in the cloud and you get the capability, but your prompts, your code, and your files travel to someone else’s servers, and every extra round of work adds to the bill.
Edge0 says that compromise just got much smaller. On November 13 the company open-sourced Edge0-145B and Edge0-120B, together with its unified inference framework and deployment tools. It calls Edge0-145B the world’s first open-source on-device AI model with 145 billion parameters, and says it brings coding, reasoning, and agent capabilities that rival specific frontier cloud models to a personal computer. Those are the company’s claims, from its own announcement and its own evaluations. No independent verification had been published when the release appeared, so read the numbers below as strong, checkable claims, not as settled results.
What actually shipped
This is a full stack release, not just a weights file. Edge0 released the model weights, the inference framework that makes the weights usable on a laptop, and the deployment tools around them, so developers can reproduce the results, evaluate or fine-tune the models, and integrate them into their own products.
The headline machine is modest by data centre standards and demanding by laptop standards. Edge0 reports:
- Memory: 29GB peak memory, which fits on a compatible 48GB laptop such as a MacBook Pro, without a server.
- Prompt processing: More than 1,100 tokens per second of prefill, so long prompts, including dozens of pages of code, are read quickly before generation starts.
- Generation: More than 80 tokens per second of decoding, which is faster than a person reads.
- Power: About 120W typical power in the reported setup, with sustained operation and no throttling.
- Context: 128K to 256K tokens, with essentially the same speed on long inputs.
Context matters here. Before September 2026, memory limits kept most on-device models below 10 billion parameters. Apple Intelligence’s on-device model has about 3 billion parameters, and Google’s Gemini Nano came in 1.8 billion and 3.25 billion versions. Edge0-145B is roughly fourteen times that old ceiling, and about four times Edge0’s own earlier phone release, Edge0-35B, which reported peak memory below 2.5GB and more than 37 tokens per second on an iPhone 18 Pro Max.
How 145 billion parameters fit in 29 gigabytes
The trick is not compression after the fact. Edge0 says it designed the model and the inference engine together, across four layers, so local execution was the starting point rather than a shrinking step applied to a cloud model.
First, the architecture shares large matrices across sparse experts. Each group of 64 experts shares two large matrices, while each expert keeps a small set of parameters that captures its differences. Previously every expert stored a complete set of matrices. The result is that of the model’s 145 billion total parameters, only about 10 billion take part in computation on each inference step.
Second, training built for local execution from day one, with deployment constraints inside the design. Third, loading pipelines across three storage tiers at once, so reading and computation overlap instead of waiting in line. Frequently used hot experts stay in GPU memory. An expert library of about 40GB sits in system memory, with experts brought into GPU memory as needed. About 48GB of vocabulary data sits on the SSD and is read on demand.
Fourth, scheduling uses the CPU and the GPU together. The GPU computes cached experts while the CPU handles the remaining work in parallel, and speculative decoding verifies multiple candidate tokens at once, spreading the cost of reading weights across more output tokens. That combination is what Edge0 credits for the reported decode rate.
None of this makes the footprint small in absolute terms. A 29GB peak, plus about 40GB in system memory for the expert library and about 48GB on disk for vocabulary data, still restricts this to high-end machines. What it changes is the class of machine required, from a server rack to a single well-specified laptop.
The benchmark claims, in full
Edge0 reports that Edge0-145B is competitive with specific frontier cloud models across coding, reasoning, knowledge, and instruction following in its own evaluations:
- Coding: 64.6 on DeepSWE 1.1, ahead of Qwen3.8-Flash-Next at 58.7 and DeepSeek-V4-Flash-0731 at 54.4.
- Agent work in a terminal: 86.52 on Terminal-Bench 2.1, ahead of Qwen3.8-Flash-Next at 83.3, DeepSeek-V4-Flash-0731 at 82.7, and Claude-Opus-4.6 (Max) at 78.2.
- Reasoning and knowledge: 90.4 on GPQA-Diamond, close to DeepSeek-V4-Flash-0731 at 90.8 and Claude-Opus-4.6 (Max) at 91.3, and 35.4 on HLE, close to Qwen3.8-Flash-Next at 35.9.
- Instruction following and long context: 80.6 on IFBench, ahead of DeepSeek-V4-Flash-0731 at 79.2 and Claude-Opus-4.6 (Max) at 62.5, and 79.3 on AA-LCR against 78.0 for Claude.
Averaged across six benchmarks, Edge0 puts Edge0-145B at a mean score of 72.8 with 29GB of measured peak memory, against 71.7 for Qwen3.8-Flash-Next with an estimated 70GB footprint for standard 4-bit deployment. In other words, comparable or higher capability at about 40 percent of the estimated memory. The smaller Edge0-120B scores lower on most measures, with a mean of 68.9 at the same 29GB peak.
Two cautions belong next to every number in this section. The Edge0, Qwen, DeepSeek, and Claude comparisons come from Edge0’s internal evaluations, and the baseline memory figures marked as estimates remain to be measured. Independent groups have not yet published their own runs. The release includes the setup and reproduction instructions, which is the right way to make a claim testable, but tested by others is a different status from tested by the maker.
Why local agents change the economics
The most practical claim is not a benchmark. It is the workflow. Edge0 says a single natural-language request can start an entire local workflow: the model breaks down the requirements, plans the steps, writes code, calls tools to run it, debugs and repairs failures, takes screenshots to inspect the result, and keeps iterating until the work is complete. The company shows this pattern applied to a Tesla Roadster explainer video produced within an hour, a Manhattan city scene built in Blender within an hour, and playable games built locally.
Whether or not every demo generalises, the cost structure is real. Agent work is expensive in the cloud because it loops. Dozens of rounds of planning, execution, and visual review each generate billed tokens. Local inference removes those token and API charges. Hardware and electricity still cost money, and 120W sustained is not free, but the marginal cost of one more iteration falls close to zero. Code and data stay on the computer, and core capabilities work offline.
That matters most in three places. Teams with sensitive code or regulated data get frontier-class help without sending the material out. Developers with unreliable connectivity get a model that does not depend on it. Heavy agent users get predictable costs, because the bill is the laptop they already own, not a meter that runs faster the more ambitious the task gets.
What to do if you want to try it
For teams considering Edge0 this week:
- Check your hardware first. You need a compatible high-end laptop, realistically a 48GB machine or better, plus disk room for the expert library and vocabulary data. This is not a model for an ordinary office laptop.
- Reproduce before you trust. Run DeepSWE or Terminal-Bench tasks from your own workload, on your own machine, and compare the output and the speed against the cloud model you use today.
- Start with work that benefits from looping. Long debugging sessions, multi-step coding agents, and visual iteration gain the most from near-zero marginal cost per round.
- Keep a cloud fallback for the hardest tasks. On HLE, the hardest reasoning measure in the set, Edge0-145B at 35.4 still trails Claude-Opus-4.6 (Max) at 40.0 in Edge0’s own table. Local first does not have to mean local only.
- Watch for independent results. The weights and framework are out, so outside evaluations should follow quickly. That is when these claims become findings.
The bottom line
Edge0-145B does not prove that the cloud era is over, and its numbers need outside checking before anyone builds a strategy on them. What it does show, if the reported 29GB peak and 80-plus tokens per second hold up on other laptops, is that the boundary between a local model and a frontier model just moved by an order of magnitude, from under 10 billion parameters to 145 billion, in open source, with the full stack included. For years the choice was capability or control. This is the strongest attempt yet to offer both on the same machine, and it is now downloadable for anyone with the hardware to test the claim.