Introducing Edge0-145B and Edge0-120B, edge0(official)’s first device-native 100B-class foundation models with open weights. Cloud-level intelligence on your laptop. Free, offline, and fast enough for real vibe coding. - Smart: beats Claude Opus 4.6 on Terminal-Bench 2.1 (86.5% vs. 78.2%) - Light: needs just 29G memory, 1/3 of a standard... setup (100G+) - Fast: decodes 75+ toks/s, prefills 1050+ toks/s on RTX Spark and Mac. - Steady: handles 128K–256K tokens of context without slowing down, even on long tasks Frontier-class agentic coding. No cloud API. No token bill. Your code and data never leave your machine. Open weights are live on Hugging Face. Run it locally or plug it into your favorite coding agent.show more

@Edge0_AI Congrats on that launch! Open models all the way :)

For 18 months, every lab’s on-device models stayed under 40B. Today, we pushed on-device AI to 145B parameters.

Frontier-level agentic coding, fully on-device. Across six major benchmarks, Edge0-145B performs on par with leading open cloud models such as DeepSeek-V4-Flash. Its edge is agentic coding: - 64.6 on DeepSWE 1.1 - 86.5 on Terminal-Bench 2.1 It’s ahead of comparable open models on both. On Terminal-Bench, IFBench, and AA-LCR, it also outscores the closed cloud model Claude Opus 4.6 (Max).

Light and fast: 100B-class models fit on a laptop. - Edge0-145B: 29GB peak memory; 72.8 avg across 6 benchmarks - Qwen3.8-Flash-Next: ~70GB; 71.7 - DeepSeek-V4-Flash: ~87GB; 70.1 Same intelligence as models that need 3× more memory.

Local vibe coding #1: 3D modeling One prompt. Wi-Fi off. We asked Edge0-145B to build Manhattan in Blender. It wrote the Python code, generated thousands of buildings on a street grid, added the skyline and rivers on both sides, and rendered a 20-second flythrough. It even checked its own renders, caught mistakes, and went back to fix them. Cloud tokens: 0. API cost: $0.

Local vibe coding #2: game worlds One prompt, five reference images, and no internet. We asked Edge0-145B to build a medieval city for a fantasy RPG. It built the whole thing from scratch: red-tiled roofs, a market square, a Gothic cathedral with twin towers, and even a vaulted interior with stained-glass windows. The tricky part was the camera path. We fixed it in advance, so the geometry had to actually work. One misplaced wall or column, and the camera would clip right through it. Cloud tokens: 0. API cost: $0.

Local vibe coding #3: video production Can a flying Roadster actually work? We gave Edge0-145B one prompt to find out. It worked through the physics, wrote a script, recreated the car in 3D, and put together a 94-second explainer. All on a laptop, without calling a cloud model. Cloud tokens: 0. API cost: $0.

Local vibe coding #4: PCB design This time we gave Edge0-145B a different task: design a circuit board, then animate how it gets made. It placed the components and routed both layers in KiCad, brought the board into Unreal Engine 5, set up the cameras and lighting, and rendered a fabrication and assembly animation. PCB design and 3D animation are usually two separate workflows, often handled by different people. Here, one model handled both on a single machine. And none of the schematics or layouts had to leave the device. For hardware teams working with sensitive designs, that last part matters a lot. Cloud tokens: 0. API cost: $0.

How do we run a 145B model on a laptop or consumer GPU? The short answer: we don't keep all 145B parameters on the GPU at once. Edge0-145B is a sparse MoE model with 512 experts per layer. Only 10 are activated for each token, which works out to roughly 6.6B active parameters. The key is our latent MoE architecture. Attention operates at the full width of 2560. Before a token reaches the experts, a shared projection compresses it to 1792 dimensions. Another projection expands it back afterward. This means the experts themselves can stay small, about 1.6 MB each. We keep a few thousand frequently used experts on the GPU and stream the rest from system RAM as needed. So GPU memory usage depends largely on how many experts we keep loaded, rather than the model's full parameter count. The model is also trained natively in 4-bit, with a built-in draft head to speed up decoding.

Both models are open today: the weights and the engine. ⛰️ Edge0-145B: flagship — 86.5 on Terminal-Bench 2.1 🪶 Edge0-120B: lighter version Runs on your laptop or desktop. Serves an OpenAI- and Anthropic-compatible API, so you can point your coding agent at localhost and keep working offline. 🤗 145B: 🤗 120B: 💻 Engine: 💻 App: Build something with it? Reply with what you made—we’ll share the best ones.

@Edge0_AI 145B open weights on a desktop, so no API key and no usage bill for a big coding model. What kind of desktop are we talking about though?

@Edge0_AI For a desktop computer, you’re looking at a decent gaming PC. RTX 4090 24GB is kinda the sweet spot, 16GB VRAM can run it

@Edge0_AI Local AI is getting powerful

@Edge0_AI @Presidentlin

@Edge0_AI I care most that it's open source and local, so I can look inside it and nobody can switch it off on me

@Edge0_AI Wait, running Edge0-145B locally on a laptop even without wifi is pretty impressive, curious how it works? Check it out on GitHub Grab the model weights on Hugging Face Join the Discord community

@Edge0_AI I'd bet people start fine tuning this on their own codebases pretty fast, the weights are right there on their machine

That's a big deal for freelance work, especially when handling client code and sensitive projects. keeping everything local makes a lot of sense. If you want to give it a try, explore Edge0-145B on GitHub: get the model weights on Hugging Face: and join the Discord community:

@Edge0_AI Keen to see how it handles a big repo when weights are moving between GPU memory and SSD. That's the first thing I'd test.

@Edge0_AI Who's signing up first? I'd guess the teams that can't let any code touch the cloud

@Edge0_AI 29G peak is the honest number. the one that bites later: weights and kv cache share the same memory, so the context you get is what's left after the model loads, not the spec sheet. that's the gap memoryrouter exists for.

@Edge0_AI Not having to pay for cloud coding tools every time would be a big plus, especially for students. definitely worth checking out Edge0-145B yourself.

@Edge0_AI if useful a principal engineer who built agent tooling for an ide platform could run your 145B as a local default. dm failed so tap below

@Edge0_AI thanks

@Edge0_AI the dm never landed but i'll take it

145B on a laptop still sounds wild. would be interesting to see what kind of setup makes it possible. curious to try it yourself? check out Edge0-145B on GitHub: get the model weights on Hugging Face: and join the Discord community: Excited to be working with y'all.

@Edge0_AI Edge0-145B also puts the inference engine on GitHub, and that's the part I'd read first. The memory handling across GPU, RAM and SSD lives there.

@Edge0_AI Any plans for MLX and native Apple silicon support?

@Edge0_AI what hardware does 145B actually need to run well locally? if it's a normal workstation, dropping the API bill is an easy sell

@Edge0_AI Interesting. I'll take a look at this Samuel.

@Edge0_AI Arithmetic check: 145B in 29GB is ~1.6 bits/weight, so it must be sparse MoE with offloaded experts. 75 tok/s at ~500GB/s implies roughly 13B active weights at 4-bit. Is 29GB before KV cache at 256K, and was Opus 4.6 run in the same harness and settings?

@Edge0_AI Can we get GGUF quants?

@Edge0_AI @grok quanto de vram é necessário pra rodar?

@Edge0_AI was the 75+ toks/s measured with the 256K context actually filled, or on a short prompt? agent loops spend most of their time deep in context

@Edge0_AI Which build produced the 86.5%, the 29GB one people will download or a larger one? In my Gemma 4 tests at temperature 0, two quantizations of the same model gave word-for-word identical answers on only 23 to 32% of prompts, so a score really belongs to one specific file.

@Edge0_AI Sounds awesome! Can it do 3d?

@Edge0_AI @grok how is this possible? how many bpw

@Edge0_AI Any plans to make the inference engine capable of using multiple GPUs? I have two RTX cards but only 16gb ea

@Edge0_AI Finally something you could point at NDA client code, since none of it leaves your machine.

@Edge0_AI the visual check is what I'd watch. an agent loop takes a ton of screenshots, and running locally none of those get billed as cloud calls

Open weights at 29G of memory is the headline for me. A 100B class model on a laptop with no token bill is a fun place for vibe coding to land. How does it hold up deep into that 256K context?

@Edge0_AI Looks quite interesting, congrats. @ddalcu mlx-serve support plz..

@Edge0_AI device native 100b is a serving claim before its a model claim. the question is whether it fits next to everything else the laptop is already doing

@Edge0_AI self-reported 86.5% on terminal-bench, launch day, vendor's own hardware. that's a claim, not a result. run it on someone else's eval and a real laptop, then we'll talk

@Edge0_AI 86.5 on terminal-bench 2.1 with ~6.6b active params is a big number, though benchmarking against opus 4.6 instead of the current opus is convenient. want to see it on a normal laptop outside the demos