Edge0 open-sources 145B-parameter model that runs on a laptop with frontier-level coding scores

Edge0 has open-sourced two large language models, Edge0-145B and Edge0-120B, together with the inference framework and deployment tools used to run them. The company describes the larger model as the first open-source on-device AI model with 145 billion parameters, and claims it delivers coding, reasoning and agent capabilities that rival frontier cloud models on a personal computer. The announcement, which surfaced on the Hacker News front page, positions the release as an answer to the familiar trade-off in local AI: capable models need server-class hardware, while small on-device models struggle with demanding work.

What Edge0 reports the model can do

According to Edge0, Edge0-145B runs on a compatible laptop such as a MacBook Pro with a peak memory footprint of 29GB. The company reports prefill speeds above 1,100 tokens/s, decoding above 80 tokens/s, roughly 120W of typical power draw, and a 128K–256K context window whose speed holds steady on long inputs. For phones, Edge0 points to its earlier 35B model, which it says reaches 37+ tokens/s on an iPhone 18 Pro Max using about 2.5GB of memory — while cautioning that the two sets of figures come from different models and hardware.

On the agent side, Edge0 says the model can carry a task through to a finished result: planning, writing and running code, debugging failures, and visually checking outputs for projects such as 3D scenes, games, hardware programs and PCB designs. Local inference also removes per-token API fees, keeps data on the machine, and keeps core capabilities working offline, though hardware and electricity still cost money.

How a 145B model fits on a laptop

Edge0 credits four coordinated design choices:

• Architecture. Groups of 64 experts share two large matrices, while each expert keeps only a small private set of parameters that captures its differences. Of the model's 145B total parameters, roughly 10B take part in any single inference step, which cuts how much data must be stored and moved.
• Training. The model is trained in its on-device configuration from the outset, rather than being compressed or pruned from a cloud-sized model after the fact.
• Loading. Storage is spread across three tiers whose reads overlap with computation: frequently used "hot experts" stay in GPU memory, an expert library of about 40GB sits in system memory, and roughly 48GB of vocabulary data lives on the SSD and is read on demand.
• Scheduling. 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.

Vendor-reported benchmarks

Edge0's own comparisons place the model near or ahead of named frontier cloud models:

• DeepSWE 1.1, coding: 64.6, ahead of Qwen3.8-Flash-Next at 58.7 and DeepSeek-V4-Flash-0731 at 54.4.
• GPQA-Diamond, reasoning and knowledge: 90.4, close to DeepSeek-V4-Flash-0731 at 90.8 and Claude-Opus-4.6 (Max) at 91.3.
• IFBench, instruction following: 80.6, ahead of DeepSeek-V4-Flash-0731 at 79.2 and Claude-Opus-4.6 (Max) at 62.5.
• Terminal-Bench 2.1: 86.52, versus 83.3 for Qwen3.8-Flash-Next, 82.7 for DeepSeek-V4-Flash-0731 and 78.2 for Claude.
• AA-LCR: 79.3, versus Claude's 78.0.

These numbers come from Edge0's announcement; independent verification has not yet been published.

Why on-device models stayed small

Edge0 frames the release against a history of constrained local models. According to the company, memory limits kept most on-device models below 10B parameters before September 2026: Apple Intelligence's on-device foundation model has roughly 3B parameters, Google's Gemini Nano ships in 1.8B and 3.25B versions, and even a 7B model could consume about 4GB of memory. Larger models hit a memory wall, since weights compete with the operating system and other apps for capacity and bandwidth. Edge0 also notes that Apple has acknowledged its on-device model was not designed for agentic applications requiring broad world knowledge.

Why it matters

If Edge0's figures hold up under independent testing, the release undercuts a long-standing assumption: that frontier-level coding and agent work requires cloud GPUs and API billing. Releasing the full stack — weights, inference framework and deployment tools — lets developers reproduce the results, evaluate or fine-tune the models, and embed them in their own products, which is a different proposition from a closed local app. The practical caveats are real: 29GB of peak memory and tens of gigabytes on disk restrict this to high-end machines, and every performance claim so far comes from the vendor. But a 145B-parameter model generating text at reading speed on a laptop, with agent workflows that plan, execute and verify code locally, would shift meaningful AI work off the cloud entirely.