Edge0 releases a 145B on-device model claiming 29GB peak memory, while HN calls it a rebranded Qwen fine-tune on a Strata fork
- Edge0 released Edge0-145B and Edge0-120B with weights, an inference framework, and deployment tools, claiming a 145B model runs on a laptop at 29GB peak memory, 1,100+ tokens/s prefill, 80+ tokens/s decode, and about 120W typical power.
- The architecture shares two large matrices across each group of 64 experts while each expert keeps a small set of its own parameters, so only about 10B of the 145B parameters take part in each inference step, with 128K to 256K context.
- Hacker News readers point out the Hugging Face page lists Qwen3.8-Flash-Next as the base model and that the edge0-strata engine is a fork of the Strata project, reading the release as a fine-tune with heavy branding.
- Commenters dispute the open source label, say the license is too restrictive even for open weights, and note the NVFP4 build is Blackwell-only with no Mac runtime despite MacBook and M2 Ultra benchmarks.
- A commenter who runs coding agents says the reported TerminalBench score is roughly on par with 5.6 Terra max, which they describe as capable of driving a cloud harness through complex knowledge work.
Hacker News 의견들
Pretty sure this is derived from Qwen3.8 Flash Next. The HF page lists it as the base model, and the memory comparison is against Qwen's standard 4-bit deployment, which would explain why it beats the base model: it loses less to quantization.
I'm running Strata + Qwen 3.8 Flash Next on 2x 3090 and 128GB of RAM and it's genuinely useful local AI. If this is an upgrade I'll take it.
The engine is called edge0-strata because it's a fork of strata. Someone fine-tuned Qwen, tweaked the model, forked the runtime, and put a lot of marketing and branding on top.
This isn't open source, it's open weights. And the license is restrictive enough that it doesn't even meet the usual open weights definition.
I'll believe it when third party benchmarks reproduce it. Skeptical at that model size, but I hope it's real.
The Blender examples impressed me. Tool use has come a long way when the tool is Blender. My worry is cleanup: agent output is usually messy enough that fixing it eats whatever time you saved.
TerminalBench scores track real performance for coding agents that lean on bash. This is roughly on par with 5.6 Terra max, which in my experience drives a cloud harness through pretty complex knowledge work.
They shipped NVFP4, so Blackwell only, and the engine looks CUDA 13 only. I want to see GGUF files and other inference engines pick it up. The variants link on HF is a directory of the model scattered across files.
They benchmarked an M2 Ultra and talk about MacBooks, but there's no Mac runtime today and no guidance on how to run it on a Mac at all. Maybe that's still in development.
145B in 29GB peak. Do the bits-per-parameter math and it's hard to see the whole model in RAM, though the sparse expert trick means only about 10B is active per token. Nobody mentions how much the context costs.
Their engine is a fork of strata, which has been doing this for a while. They aren't hiding it, the README says so. Still weird that the marketing page reads like it's all new work.
So did Strata versus llama.cpp, and Strata pulls 2.5 to 3x the performance of the codebase it forked. If they improved the architecture further by sharing experts, this is plausible.
What's the 120B model for?
Fitting in my 48GB MacBook once the Mac build lands. 145B looks too big for that.
Harness work matters as much as the model. Context pruning and healing malformed tool calls took Qwen 27B from 73 to 85 on Terminal Bench 2.1 for me.