Edge0's 145B Laptop Model Runs in 29GB, but Its Weights Sit on Qwen3.8-Flash-Next Under a Community License

Edge0 has published Edge0-145B and Edge0-120B, two large mixture-of-experts models it says can run on a personal laptop rather than a server. The company announced the release on November 12, 2026, together with its inference framework, deployment tools and a desktop app. Its headline claim is that a 145-billion-parameter model can run locally with 29GB of peak memory, decode at more than 80 tokens per second and prefill at more than 1,100 tokens per second, at about 120W of typical power in the reported setup.

The weights are on Hugging Face as NVFP4 and q3k builds for both sizes, and the framework is on GitHub. The licence and lineage are more specific than the “open-source” framing in the launch post suggests, and they matter for anyone planning to ship the model.

How a 145B model fits in 29GB

Edge0 says only about 10B of the 145B parameters take part in each inference step. The bigger change is in how experts are stored. Each group of 64 experts shares two large matrices, while each expert keeps a small set of its own parameters that capture how it differs. That cuts the weight data that has to be stored and moved without reducing the expert count.

The runtime spreads the model across three tiers. Frequently used “hot experts” stay in GPU memory. An expert library of roughly 40GB sits in system memory and is pulled into the GPU as needed. About 48GB of vocabulary data stays on the SSD and is read on demand. The GPU computes cached experts while the CPU handles the rest in parallel, and speculative decoding checks several candidate tokens at once so each weight read produces more output. Edge0 says context runs from 128K to 256K tokens at roughly the same speed.

Day-one support covers macOS, Windows and Linux, including Apple M-series chips and NVIDIA platforms such as RTX Spark, RTX 4090 and RTX 5090. The company says local use works on a 48GB laptop.

Benchmarks are Edge0’s own

Edge0 reports 64.6 on DeepSWE 1.1, 86.52 on Terminal-Bench 2.1, 35.4 on HLE, 90.4 on GPQA-Diamond, 80.6 on IFBench and 79.3 on AA-LCR for the 145B model. Its comparison table puts Qwen3.8-Flash-Next at 58.7 on DeepSWE and 83.3 on Terminal-Bench, and DeepSeek-V4-Flash-0731 at 54.4 and 82.7. The smaller Edge0-120B scores 53.6 on DeepSWE and 81.3 on Terminal-Bench.

The company says the Edge0, Qwen, DeepSeek and Claude results come from its internal evaluations, and that full setup and reproduction instructions will ship with a technical report. Across the six benchmarks it gives Edge0-145B a mean of 72.8 against 71.7 for Qwen3.8-Flash-Next, whose roughly 70GB footprint for standard 4-bit deployment is an Edge0 estimate rather than a measurement.

Built on Qwen, licensed in two layers

The launch post says Edge0 trains the model “in its on-device architecture from the outset.” The Hugging Face model cards, however, list Qwen/Qwen3.8-Flash-Next as the base model, and the README thanks the Qwen3.8-Flash-Next and Strata teams, noting that Edge0 extends the Strata engine with latent MoE support.

The weights are not under Apache 2.0, the licence Edge0 used for earlier releases such as Edge0-35B-A3B-preview. The new repositories carry the Edge0 Community Model License 1.0, dated November 12, 2026. It covers only the Edge0-specific contributions listed in a release-scope file, such as model modifications, adapters and training deltas. Qwen materials stay under the Qwen Community License 1.0, reproduced unchanged in the repository, and the NOTICE file says Edge0 permission does not replace any separate Qwen authorisation.

According to the model card, the Edge0 licence permits personal use, research, modification, noncommercial sharing, internal use and commercial use of generated outputs. Selling or commercially bundling the model, running a commercial model service, or offering an independent commercial coding or office assistant built on it needs a separate written licence from Edge0. The inference framework remains Apache 2.0, and the bundled Strata fork keeps its MIT licence.

Where it fits in Edge0’s line-up

Edge0-145B follows Edge0-35B, which the company released on September 10 for phones and says runs at 37+ tokens per second in about 2.5GB on an iPhone 18 Pro Max. Edge0 now positions its 145B and 120B models for laptops and vibe coding, and its 35B and 8B models for phone-based agent tasks.

The company is clear that the cloud still has a role. Very low-compute devices cannot run a 145B model, and long-term memory across devices or searches that need current information still require online services. Edge0’s pitch is that more demanding work, including multi-step coding projects, can now stay on the machine and keep running offline.