Edge0-145B’s weights are out. It runs on one NVIDIA GPU, but not yet on a Mac

What Edge0 released

On November 12, 2026, Edge0 announced that it was “open-sourcing Edge0-145B and Edge0-120B, together with our unified inference framework and deployment tools.” Its GitHub page dates the release November 13. Edge0 calls the larger model “the world’s first open-source on-device AI model with 145 billion parameters,” with “coding, reasoning, and agent capabilities that rival frontier cloud models.” They’re the project’s largest models so far: in September it released a 35B model that runs on phones and an 8B checkpoint, and its repository has passed 10,000 stars on GitHub.

Edge0-145B isn’t trained from scratch. Its model page calls both models “derived from Qwen3.8-Flash-Next,” Alibaba’s open model, which Qwen’s card lists at 125 billion language-model parameters with 6 billion active. Edge0 rebuilt it as what it calls a latent mixture of experts, 145 billion parameters in all, with 10.12 billion active for each token: 512 experts in each of 48 layers, of which 10, plus one shared expert, do the work. The experts read compressed inputs, which is what lets the engine keep a few thousand of them on the graphics card and stream the rest from system memory. It has a 256K-token context, and the server starts at 131,072 tokens unless you ask for more.

Edge0-120B is the same design with smaller experts, at 8.64 billion active parameters. Each model comes in two builds, Q3_K and NVFP4, which Edge0 says were “trained natively at 4 bits.”

How good it is

These are Edge0’s numbers, from its model page. Edge0 says its own, Qwen’s, DeepSeek’s and Claude’s results “come from Edge0’s internal evaluations,” and that “the complete evaluation setup, hardware environment, and reproduction instructions” will come with its technical report. Higher is better.

Benchmark	Edge0-145B	Qwen3.8-Flash-Next	DeepSeek-V4-Flash-0731	Qwen3.8-27B	Claude Opus 4.6 (Max)
DeepSWE 1.1	64.6	58.7	54.4	42.2	No score
Terminal-Bench 2.1	86.5	83.3	82.7	73.0	78.2
AA-LCR	79.3	79.1	79.7	82.0	78.0
IFBench	80.6	81.3	79.2	79.5	62.5
GPQA-Diamond	90.4	91.7	90.8	89.2	91.3
HLE	35.4	35.9	33.8	30.8	40.0

So the claim is about coding. On DeepSWE 1.1 and Terminal-Bench 2.1, Edge0-145B leads the table, including the Qwen model it was built from; on the other four tests it trails at least one rival. The Claude in the table is Opus 4.6, several releases behind Anthropic’s current Opus 5.5. And Edge0’s two pages disagree about the smaller model: the model page gives Edge0-120B 51.3 on DeepSWE 1.1 and 75.3 on Terminal-Bench 2.1, and the announcement gives it 53.6 and 81.3. We found no independent evaluation by November 13.

What it takes to run

Edge0’s requirements are “Linux or Windows with an NVIDIA GPU and CUDA 13,” with “at least 16 GiB of GPU memory and 20 GiB of system memory.” Its Windows release asks for “compute capability 7.5 or newer (RTX 20 / 30 / 40 / 50)” on x64 PCs, and the NVIDIA CUDA Toolkit 13 installed. The download is large: on Hugging Face, the Q3_K folder adds up to about 98 GiB and the NVFP4 folder about 105 GiB. Almost 48 GiB of either is an embedding table that, by default, the engine reads from disk as it needs it.

Here are Edge0’s measured speeds for Edge0-145B, in tokens a second: how fast it reads a 16,000-token prompt, and how fast it writes replies in chat and in agent work.

Machine	Build	16K-token prompt	Chat replies	Agent replies
RTX 5090 D, 24 GB	Q3_K	3554.9	99.7	121.2
RTX 5090 D, 24 GB	NVFP4	3165.7	79.6	98.3
RTX 4090, 24 GB	Q3_K	3179.4	69.9	91.7
RTX Spark N1X, 60 GiB	NVFP4	922.4	40.1	46.0
RTX 5090 D held to 16 GiB VRAM and 20 GiB RAM	Q3_K	335	36.8	44.9
Apple M2 Ultra	Q3_K	800.3	60.9	57.0

The fifth row is the same card held to the 16 GiB and 20 GiB in Edge0’s requirements, a budget its README calls “a test limit, not a minimum system requirement.” Without those limits, the same card held the Q3_K build as 23.14 GiB of GPU memory plus 31.28 GiB of system memory.

The Mac is the catch. The announcement says “day-one support covers macOS, Windows, and Linux, including Apple M-series chips,” and that the model “runs on a compatible personal laptop, such as a MacBook Pro.” The release doesn’t bear that out. The 0.2.0 release is “Windows only,” the README’s platform table leaves macOS blank for the two big models, the edge0-strata guide lists its Apple Silicon engine among the parts removed, and the model page says its M2 Ultra measurements “do not establish macOS release availability.” Commenters on Hacker News noticed the same, and one questioned the announcement’s “29GB peak memory” headline, since the same post puts “an expert library of approximately 40GB” in system memory and “approximately 48GB of vocabulary data” on the SSD. The smaller Edge0 35B is the model that runs on a Mac today.

How to run it

There’s no hosted version to try: Hugging Face lists no inference provider for it, and OpenRouter doesn’t carry it. Hugging Face’s page also shows its usual one-line commands for Ollama, llama.cpp, LM Studio and vLLM, but Edge0’s docs run the model only through its own engine, and a model folder is a set of Edge0’s own files, such as experts.bin and a “skeleton” GGUF, rather than one GGUF those apps load. Its own route has three parts:

1. Build edge0-strata, the engine for these two models, as its guide describes. On Linux that takes “CUDA 13, CMake 3.24 or newer, Ninja, a C++17 compiler, Python 3.10 or newer, and uv.”
2. Build edge0-infer, the program you run, from the edge0-infer folder of the repository with cmake --preset cpu && cmake --build --preset cpu. On Windows, the 0.2.0 release also has a prebuilt Edge0 Chat app, which the README says opens a model folder such as Edge0-145B-q3k; it still needs the CUDA Toolkit.
3. Download one model folder from Hugging Face, such as Edge0-145B-q3k, and keep its files as they are. edge0-infer identifies the model from the folder’s model.json.

Then check the folder and start the server:

export EDGE0_145B=/path/to/Edge0-145B-q3k
build/cpu/bin/edge0-infer inspect --model "$EDGE0_145B" --json
build/cpu/bin/edge0-infer serve --model "$EDGE0_145B"

By default, serve waits up to 1,800 seconds for the engine to load the model. When the server prints its Listening: line, it answers at http://127.0.0.1:8000/v1 under the model name edge0_145b, with OpenAI’s Chat Completions and Responses formats and Anthropic’s Messages format, as its server guide lists. To chat in the terminal instead, run edge0-infer chat --model "$EDGE0_145B", but not at the same time as serve on the same GPU.

If your NVIDIA card is in a PC and you work on a Mac, the server can listen on your network with --host. The guide says a listener beyond your own computer requires --api-key-file, and clients elsewhere “must supply its token.”

Give it real work

Edge0 doesn’t document a coding agent, but OpenCode takes any server that speaks OpenAI’s API. Add a provider to opencode.json, as OpenCode’s docs describe, with the key under models matching the name the server uses:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "edge0": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Edge0 (local)",
      "options": { "baseURL": "http://127.0.0.1:8000/v1" },
      "models": { "edge0_145b": { "name": "Edge0-145B (local)" } }
    }
  }
}

Two of the model’s known limitations matter for agents. “Concurrent requests wait in edge0-strata’s queue,” so an agent that sends several at once will wait for each, and “identical requests at temperature 0 can produce different answers.” Start with a small task, and watch its tool calls before you leave it alone.

Enjoy runs OpenCode beside Claude Code, Codex and the other labs’ agents, and its model picker lists every model OpenCode has set up, local ones included. So when a plan runs low, you can switch the conversation to a local model and keep going. Enjoy doesn’t run models itself, so Edge0-145B still needs the NVIDIA machine above.

What the license allows

The model weights come under Edge0’s own Community Model License 1.0, which has commercial limits, unlike the Apache 2.0 license on Edge0’s engine code. It allows “personal use, including local deployment,” “internal organizational use, including use by businesses,” paid client work, and “commercial use of Outputs.” Three things need “a separate written commercial license from Edge0”: selling the model or paid access to it, running a paid inference service on it, and selling “an independent product primarily designed for AI-assisted coding or office productivity” built on it, “whether inference is hosted or runs on a user’s device.”

Qwen’s license still applies on top, as Edge0’s notice says, and Edge0’s license adds: “Permission from Edge0 is not permission from Qwen.” Qwen’s text asks companies that run a model service or an AI coding or office assistant business to get a separate license from Qwen for commercial use.

What isn’t known

• When it comes to the Mac. Edge0 hasn’t given a date, or explained why its announcement says the Mac is supported today.
• Independent scores. Every number above is Edge0’s, and its technical report isn’t out.
• Where the 29 GB figure comes from. The announcement doesn’t name the machine or settings behind it.
• Which Edge0-120B scores are right. Its two pages differ.

If you have an NVIDIA card with 24 GB and the RAM to go with it, Edge0-145B is worth a download to test against your own work, remembering that every score so far is Edge0’s. On a Mac, wait for Edge0 to ship the support its announcement describes.