Edge0 open-sources 145-billion-parameter AI model built to run on a laptop

Edge0 has open-sourced Edge0-145B, a 145-billion-parameter language model the company says was designed from the start to run on personal computers rather than in the cloud. Edge0 released the model alongside a smaller sibling, Edge0-120B, plus the inference framework and deployment tools used to run them, according to the company’s announcement.

Edge0 says Edge0-145B runs on a compatible laptop, such as a MacBook Pro, with reported peak memory use of 29GB and decoding speeds above 80 tokens per second. That marks a jump from Edge0’s earlier phone release, Edge0-35B, which the company says reached more than 37 tokens per second using roughly 2.5GB of memory on an iPhone 18 Pro Max. Edge0 notes the two figures come from different models and hardware and should not be compared directly.

Edge0 attributes the gains to changes across the model’s architecture, training, memory loading and task scheduling. Groups of 64 experts, part of a mixture-of-experts design where only some sub-networks fire per token, now share two large weight matrices instead of each expert storing its own full set, the company says. As a result, only about 10 billion of the model’s 145 billion total parameters take part in computing any single output. Edge0 also says it trained the model for local deployment from the outset, rather than shrinking a cloud model afterward, and that its inference engine splits memory loading across GPU, system memory and SSD storage while running CPU and GPU work in parallel.

On benchmarks Edge0 reports from internal testing, Edge0-145B scored 64.6 on the coding test DeepSWE 1.1, ahead of Qwen3.8-Flash-Next’s 58.7 and DeepSeek-V4-Flash-0731’s 54.4. On the reasoning benchmark GPQA-Diamond, it scored 90.4, just below DeepSeek-V4-Flash-0731’s 90.8 and Claude-Opus-4.6’s 91.3. On IFBench, which measures instruction-following, Edge0-145B scored 80.6, ahead of both DeepSeek-V4-Flash-0731’s 79.2 and Claude-Opus-4.6’s 62.5. Edge0 says it averages 72.8 across six benchmarks while using a measured 29GB of memory, against an estimated 71.7 average and roughly 70GB footprint for Qwen3.8-Flash-Next under standard 4-bit deployment, about 40% of that estimated memory for comparable or better scores, by Edge0’s own accounting.

What this means

If Edge0’s figures hold up under independent testing, the release would challenge the assumption that on-device models must trade capability for size, a limit that kept most phone and laptop models under 10 billion parameters and pushed Apple and Google toward on-device systems in the low single digits of billions of parameters. Running a 145-billion-parameter model locally would also cut the recurring API fees and connectivity dependence of cloud assistants, since Edge0 says core functions keep working offline, though hardware and electricity costs still apply.

What remains unverified is how these numbers hold up outside Edge0’s own testing, since all the comparisons above come from the company’s internal evaluations rather than an outside lab. Edge0 says a technical report with its full evaluation setup and reproduction instructions is still to come.

Edge0 says the complete model weights, inference framework and deployment tools are available now for developers to reproduce, evaluate, fine-tune or build on.