Edge0-145B: what it is, what it needs, and how to run it (2026)

What is Edge0-145B?

Edge0-145B is an open-weight language model from Edge0, announced on November 12, 2026 alongside a smaller sibling, Edge0-120B. Edge0's launch post pitches it as the first 145-billion-parameter model built to run on a personal computer, with coding, reasoning and agent skills close to cloud models.

I build the AI agents inside eesel's product, so I read launches like this for the mechanism first and the marketing second. That turned out to matter here, because the Hugging Face model card and the GitHub repo tell a more precise story than the launch post. I don't have an RTX card on my desk, so everything below comes from those docs and the release notes, not from my own run. For a verdict on quality, my colleague's Edge0-145B review digs into that separately.

Here is what Edge0 has published about the model itself:

The base model is the part the launch post buries. The model card lists Qwen/Qwen3.8-Flash-Next as the base, and the acknowledgements credit both Qwen and the open Strata inference engine. That's normal practice in open weights, and it's useful to know: if you already run Qwen3.8-Flash-Next, Edge0-145B is a re-architected, retrained version of a model you know, not an unrelated family.

How does a 145B model fit on one GPU?

It doesn't fit, not all at once. The trick is that only about 10B of the 145B parameters work on any given token, so the engine only needs those close to the GPU. Edge0 combines four ideas to make that work:

1. Shared expert matrices. In a normal mixture-of-experts model every expert stores its own full set of weights. Edge0 projects each token down to a smaller "latent" width, and the large projection matrices are shared across experts. Each expert keeps only a small record (about 1.6 MB), so thousands of them can sit in GPU memory.
2. Three storage tiers. Frequently used experts live on the GPU. The rest sit in system RAM. A 51.2B-parameter lookup table that the model reads sparingly stays on the SSD and is read on demand.
3. CPU and GPU together. The GPU computes the experts it already holds while the CPU handles the others in parallel.
4. Speculative decoding. A small 1.26B draft head proposes several tokens, and the main model checks them in one pass, which spreads the cost of reading weights across more output.

Edge0 also trained the model in this shape from the start rather than squeezing a cloud model after the fact. The architecture explains both the good news (it runs on a gaming card) and the catch (it leans hard on system RAM and SSD speed), which the next section is about.

What hardware do you actually need?

This is where the launch post and the release files disagree. The launch post's headline is "29GB peak memory" on "a compatible personal laptop, such as a MacBook Pro". The files give a fuller picture.

The v0.2.0 release notes set the minimum for Windows x64: an NVIDIA GPU of compute capability 7.5 or newer (RTX 20, 30, 40 or 50 series), at least 16 GiB of GPU memory, 20 GiB of system memory, and the CUDA Toolkit 13. On Windows Arm64 the target is the RTX Spark N1X.

At full speed it uses more. The model card says that with residency enabled, the NVFP4 build used 23.25 GiB of VRAM plus 38.27 GiB of host RAM on an RTX 5090 D 24 GB, and the Q3_K build used 23.14 GiB plus 31.28 GiB. The 16 GiB plus 20 GiB numbers come from the same card with limits applied, at a lower speed.

The Mac claim is the bigger gap. The GitHub README's platform table marks macOS as unsupported for the 145B and 120B models, and the model card says its Apple M2 Ultra benchmark "does not establish macOS release availability". Readers on Hacker News caught it on day one:

"The blog post mentions Mac repeatedly, saying things like 'day-one support covers macOS'. But the released stuff so far seems to be CUDA only."

phillennium, Hacker News

So if you're on Apple Silicon, the smaller Edge0-35B is the Edge0 model that runs today. For anything bigger on a Mac, a hosted model like Qwen 3.8 Max or DeepSeek V4.1 Flash is the practical route for now.

Pick your machine to see where you land:

How fast is Edge0-145B?

Fast on the right card. The model card's speed table, all from Edge0's own measurements, in tokens per second:

"Agent decode" runs faster than chat because agent output (code, tool calls) is more predictable, so the draft head's guesses land more often. Two practical notes: the Q3_K build is smaller and faster than NVFP4 on the same card, and the RTX 5090 D is a China-market variant of the 5090, so a standard 5090 should be at least as quick.

How good is it, according to Edge0?

On Edge0's internal evaluations, Edge0-145B leads its Qwen base on the two agent and coding tests and trails it slightly on knowledge:

Three caveats before you lean on those numbers. First, Edge0 ran these itself, and a full technical report with reproduction steps is promised but not out yet. Second, the cloud comparison is Claude Opus 4.6, which is two generations behind the current Claude Opus 5.5 and Claude Fable 5.2. Third, Edge0's own pages disagree on the 120B: the launch post gives it 53.6 on DeepSWE and 81.3 on Terminal-Bench, while the GitHub README says 51.3 and 75.3.

Terminal-Bench is still a useful signal for agent work, as one developer pointed out:

"TerminalBench scores often correspond to real world performance for coding agents that rely heavily on bash."

rybosome, Hacker News

If you want a broader way to judge models on your own tasks, my notes on benchmarking AI support agents and the LLM tracking tools roundup cover the method.

How do you install and run Edge0-145B?

There are two routes, and both need the NVIDIA setup above.

The desktop app (Windows only). Download edge0-chat-0.2.0-windows-x64.zip from the releases page, download a model folder from Hugging Face, open Edge0 Chat and pick the folder. Edge0 says the app's core features work offline with no account or usage quota.

The command line (Linux or Windows). Clone the repo, build the edge0-strata engine and edge0-infer from the instructions in the repo, then point it at one model folder:

git clone https://github.com/Edge0-AI/Edge0.git edge0
build/cpu/bin/edge0-infer serve --model /path/to/Edge0-145B-Q3_K

Once the weights load, it prints Listening: http://127.0.0.1:8000 and serves OpenAI- and Anthropic-compatible endpoints under the model name edge0_145b. edge0-infer chat opens an interactive session instead.

Two settings matter most if memory is tight. --mmap-experts reads uncached experts from the SSD instead of holding them in RAM, and --resident-budget-gib 20 caps the RAM expert cache at 20 GiB, which needs a static expert profile file. Both live in the model folder's model.json so the app and the CLI pick them up.

Using it from a coding agent

Because the local server speaks the Anthropic Messages format, tools that accept a custom base URL can, in principle, point at it. For Claude Code, that means setting ANTHROPIC_BASE_URL to http://127.0.0.1:8000 and choosing edge0_145b as the model, the same pattern as for any custom endpoint. My Claude Code environment variables and model selection guides cover the variables.

OpenClaw and other OpenAI-compatible tools can use the /v1/chat/completions endpoint. I haven't tested either path myself, so treat this as the documented route rather than a verified one.

What does the license allow?

The launch post calls Edge0-145B "open-source". The license is narrower than that word suggests, a point raised within minutes on Hacker News:

"This is not open source, this is open weights"

Zetaphor, Hacker News

Per the model card, the Edge0 Community Model License 1.0 allows:

• Personal use, research and modification
• Noncommercial sharing and internal use at a company
• Commercial use of the model's generated outputs

It requires a separate written license from Edge0 for:

• Selling or commercially bundling the model
• Running a commercial model service
• Offering an independent commercial coding or office assistant built on it

Qwen's community license applies on top, since the model is derived from Qwen materials. The engine code itself is Apache-2.0. In practice: fine for a developer's own machine or an internal tool, not fine as the engine of a product you sell without talking to Edge0 first. My open-source AI agents roundup lists tools with looser terms if that's a blocker.

What does Edge0-145B mean for support teams?

The appeal for support is obvious: customer data that never leaves your hardware, and no per-token bill. That's a real concern I hear about. On one eesel sales call, a technical evaluator at a hardware company wanted assurance that the AI would answer only from approved knowledge, not from web or general training data, before going any further.

But a local model answers a different question. Where the model runs decides where data is processed. What it says to a customer depends on the layer around it: which help articles and past tickets it can read, what it does when it can't find an answer, when it hands off to a person, and whether anyone tested it on real conversations first. A 145B model on an RTX card ships with none of that, and the license rules out turning it into a commercial support assistant without a deal.

That layer is the job of eesel's AI helpdesk teammate. It joins the queue you already run in Zendesk or Freshdesk, answers from your own knowledge, and gets simulated against past tickets before it goes live.

SE Ranking runs one eesel agent per product on Zendesk across three products, handling 4,650 conversations in a month from about 16,800 pages of knowledge, with refunds and requests for a person handed to the team. Innr, a Dutch smart lighting company, has eesel draft Freshdesk replies in the customer's own language from 1,500 help articles and past tickets.

If you're weighing models for your own support content, my guides on connecting Claude to your helpdesk, MCP for customer support and AI agent handoffs go further than a model card can.

Should you try Edge0-145B?

Yes, if you have an RTX card with 16 GiB or more, 32 GiB or more of RAM and a fast SSD, and you want a strong local coding model with no token bill. Use the Q3_K build, run it in Claude Code or your usual agent on a task you've already solved, and compare against your cloud model.

Wait, if you're on a Mac, on AMD, or planning to ship it inside a commercial product. The first two need platform support that isn't released, and the third needs a license conversation. In the meantime, Qwen3.8-Flash-Next is the base to watch, and my Qwen overview and Kimi K3 write-ups cover the nearest open alternatives.

Try eesel for AI on your support queue

If the reason you looked at Edge0-145B was AI that answers your customers, eesel gets you there without buying a GPU. It's an AI helpdesk teammate that plugs into Zendesk, Freshdesk and other helpdesks, answers only from your help center and past tickets, and is tested on historical conversations before it touches a live one, the way SE Ranking runs it across three products. See the pricing or try eesel free.