Edge0-145B is a 145-billion-parameter model built to run on one consumer graphics card, and the engineering underneath is real. On Edge0's own tests it scores 86.5 on Terminal-Bench 2.1 and 64.6 on DeepSWE 1.1, ahead of its base model, Qwen3.8-Flash-Next.
The launch claims need footnotes, though. The "29GB" figure leaves out about 31 GiB of system RAM, macOS support isn't in the release, and the weights carry a license that restricts commercial services. My verdict: 3/5, worth testing on an NVIDIA box, not worth rebuilding around yet.
For support teams chasing privacy, the model matters less than what the bot may read. eesel answers only from your approved knowledge, and with the eesel CLI any coding agent can upload that knowledge and run test questions before anything goes live.
How I reviewed Edge0-145B
Edge0 published the release on 13 November, so this review is two days in. I build integrations and APIs at eesel, so I read releases like this the way I read any vendor's docs: the launch post tells you what they want you to believe, and the README tells you what actually ships. I worked from three sources, in this order of trust:
- Edge0's own release files. The GitHub README, the Hugging Face model card, and the license provenance file in the repo. These carry the measured numbers.
- Edge0's launch post. Useful for intent and demos, but where it disagrees with the README, I side with the README.
- Early users. The Hacker News launch thread passed 330 points and 100 comments, with a few people trying it on their own hardware.
I didn't run my own benchmark suite, and no independent lab such as Artificial Analysis had a page for Edge0-145B on 14 November. Treat every score below as Edge0's. If you want the broader model landscape first, my colleagues' DeepSeek V4 Flash and Qwen 4 Max write-ups cover the two families Edge0 compares itself against.
What is Edge0-145B?
Edge0-145B is a latent mixture-of-experts model from Edge0, a team that previously shipped the Edge0-35B phone model in September and the Audio8 speech models. It has 145B total parameters but only about 10.12B active per token, which is how it fits on hardware that could never hold a dense 145B model.
Here is what the model card says about it:
| Spec | Edge0-145B | Edge0-120B |
|---|---|---|
| Total / active parameters | 145B / 10.12B | 120B / not stated |
| Layers and experts | 48 layers, 512 experts, 10 routed + 1 shared per token | Same |
| Expert latent width | 1,792 | 1,280 |
| Context window | 256K (KV int8, 32K resident) | 256K |
| Builds on Hugging Face | Q3_K (40.81 GiB), NVFP4 (47.79 GiB) | Q3_K (27.98 GiB), NVFP4 (32.96 GiB) |
| Base model | Qwen3.8-Flash-Next | Qwen3.8-Flash-Next |
| Engine | edge0-strata (fork of Strata) | edge0-strata |
Two lines in that table carry most of the story. The base model is Qwen, and the engine is a fork of the open-source Strata project. Edge0 says so plainly in the README, though the launch post never mentions Qwen at all. That gap is what most early commenters fixed on:
"Based on Qwen3.8-Flash-Next and Strata. They made some interesting tweaks and did some fine tuning on the model, but it's interesting how much marketing speak was layered on their website to hide this fact."
Being a derivative isn't a flaw. Most good open models now start from someone else's base, and the real question is whether Edge0's changes add anything. On that, the answer looks like yes.
What Edge0-145B does well
Three things stood out, and they're all about engineering rather than raw intelligence.
The expert design is clever
Edge0 shares two large matrices across each group of 64 experts, and each expert keeps only a small set of its own parameters. The model card spells out the result: every expert is a fixed 1,630,720-byte record, so the engine can keep a few thousand of the busiest experts on the graphics card and stream the rest from system memory. The cost of that streaming tracks how many experts stay resident, not how big the model is.
Edge0 also trained the model in this shape from the start, natively at 4 bits, instead of shrinking a cloud model after the fact. Speculative decoding through a 1.26B draft head spreads the cost of each weight read across several output tokens. None of this is magic, but it's careful work, and it's why a 145B model runs at all on a 24 GB card.
Its coding and agent scores beat its base model
Edge0's benchmark table shows the largest gains where they matter for local agents:
| Model | DeepSWE 1.1 | Terminal-Bench 2.1 | HLE | GPQA-Diamond | IFBench | AA-LCR |
|---|---|---|---|---|---|---|
| Edge0-145B | 64.6 | 86.5 | 35.4 | 90.4 | 80.6 | 79.3 |
| Edge0-120B | 51.3 | 75.3 | 31.0 | 87.4 | 79.3 | 81.0 |
| Qwen3.8-Flash-Next | 58.7 | 83.3 | 35.9 | 91.7 | 81.3 | 79.1 |
| DeepSeek-V4-Flash-0731 | 54.4 | 82.7 | 33.8 | 90.8 | 79.2 | 79.7 |
| Qwen3.8-27B | 42.2 | 73.0 | 30.8 | 89.2 | 79.5 | 82.0 |
| Claude Opus 4.6 (Max) | n/a | 78.2 | 40.0 | 91.3 | 62.5 | 78.0 |
The pattern is honest in one way: Edge0-145B beats its base on the two agent benchmarks and slightly trails it on knowledge and reasoning (GPQA, HLE, IFBench). That's what you'd expect from fine-tuning aimed at coding agents. Terminal-Bench in particular tends to track how well a model drives a shell, and one early user put it in context:
"This score is roughly on par with 5.6 Terra max, which in my direct experience is perfectly capable of driving a cloud agent harness that can do pretty complex knowledge work."
If you'd rather compare against current hosted models, my colleagues' GPT-6.1 Terra review and Claude Opus 5.5 review have the independent numbers that Edge0's table lacks.
It's fast on the right hardware
On Edge0's own RTX 5090 D 24 GB machine, the Q3_K build hit 121.2 tokens per second in agent decoding and 3,554.9 tokens per second on a 16K-token prefill. That's quick enough that a local coding loop feels interactive. The launch demos (a Blender city scene, a drone program, a KiCad board) show what that speed is for: dozens of plan, run, and check rounds with no per-token bill. The Edge0 GitHub page has the full speed table and a two-minute showcase video.
Where Edge0-145B falls short
Here the launch post and the release files tell different stories. Each point below comes from Edge0's own documents.
"29GB peak memory" isn't the whole footprint
The launch post leads with 29GB peak memory and says the model runs on a 48GB laptop. The same post later describes an expert library of about 40GB in system memory and about 48GB of vocabulary data on the SSD. The README's measured numbers settle it:
With expert residency on, the Q3_K build used 23.14 GiB of GPU memory plus 31.28 GiB of system RAM. The NVFP4 build used 23.25 GiB plus 38.27 GiB. The 29GB figure looks like one slice of that, most likely the GPU side plus overhead, presented as the whole.
To be fair to Edge0, it also ran the model in a tighter budget: 16 GiB of VRAM and 20 GiB of RAM, with uncached experts read from SSD. It still worked, at 36.8 tokens per second in chat and 44.9 in agent mode. So a smaller machine can run it, just much slower. A Hacker News reader who had hoped to run it on a 48GB MacBook summed up the disappointment:
"If we're talking about ~70GB combined RAM footprint then that's not going to cut it. Still very interesting though."
macOS is in the launch post, not the release
The launch post says day-one support covers macOS, Windows, and Linux, including Apple M-series chips. The README's platform table says something narrower: the 145B and 120B models run on Linux or Windows with an NVIDIA GPU and CUDA 13, and the macOS, iOS, and Android columns are empty for them. Edge0 did publish an Apple M2 Ultra result (57.0 tokens per second agent decode) and adds that it "does not establish macOS release availability." The Edge0 Chat desktop app is Windows-only for now.
Here's what that means in practice:
Mac users can still run the older Edge0-35B on Metal. For the 145B, you need an NVIDIA box today.
The benchmarks are self-reported, and they don't fully agree
Every Edge0, Qwen, DeepSeek, and Claude score in the table comes from Edge0's internal evaluations. Edge0 says the full setup and reproduction steps are "being released with our technical report," and the only technical report published so far covers the 35B model. The numbers also shift between documents: the launch post lists Edge0-120B at 53.6 on DeepSWE and 81.3 on Terminal-Bench, while the README lists 51.3 and 75.3.
The first hands-on reports are mixed. One tester on an NVIDIA DGX Spark couldn't get a useful answer:
"It spent 8,000+ tokens at default thinking and never produced a coherent answer."
One test isn't a verdict, but it's a reason to wait for an independent scorecard before trusting the table. My colleagues' DeepSeek V4.1 Flash review shows what that looks like when a lab has run the numbers.
"Open source" comes with a commercial catch
The launch post says Edge0 is "open-sourcing" the models. The framework code is Apache 2.0, and Strata stays MIT. The weights are different. The model card puts them under the Edge0 Community Model License 1.0, with Qwen's materials still under the Qwen Community License. Personal use, research, modification, internal use, and commercial use of outputs are allowed. Selling or bundling the model, operating a commercial model service, or offering a commercial coding or office assistant needs a separate written license.
For a hobbyist or an internal tooling team, that's fine. For anyone planning to build a product on it, read the license before the README. The Hugging Face review on our blog covers how these custom licenses show up across the hub.
Edge0-145B against the alternatives
Here's how I'd place it next to what a developer would actually consider this week:
| If you want... | Pick | Why |
|---|---|---|
| The strongest local coding agent on one NVIDIA card | Edge0-145B | Best self-reported Terminal-Bench and DeepSWE scores in its class, 91.7 to 121.2 tok/s on 24 GB cards |
| The same family with fewer surprises | Qwen3.8-Flash-Next | The base model, wider runtime support, close scores on reasoning |
| A local model on a Mac today | Edge0-35B or another Metal-ready open model | Edge0-145B has no macOS runtime yet |
| Independent benchmarks and no hardware | DeepSeek V4 Flash or Kimi K3.1 Pro | Hosted, measured by third parties, low per-token prices |
| Frontier quality and you'll pay for it | Claude Opus 5.5 or GPT-6 Sol | Still ahead on the hardest reasoning |
If cost is what's pushing you local, run the math first. The Anthropic API pricing and OpenAI API pricing guides help you compare per-token bills against a GPU that also draws about 120W while it works.
My verdict on Edge0-145B
| Area | Score | Note |
|---|---|---|
| Engineering | 4.5/5 | Shared-matrix experts and SSD streaming are a real step for local inference |
| Coding and agent scores | 4/5 | Strong, but self-reported only |
| Honesty of launch claims | 2/5 | 29GB, macOS, and "open source" all need footnotes |
| Ease of setup | 2.5/5 | Build from source, NVIDIA plus CUDA 13 only, Windows-only app |
| License | 2.5/5 | Fine for internal use, restrictive for products |
Overall: 3/5. If you have an RTX 4090 or 5090 and want a local coding agent, it's worth an afternoon. If you're on a Mac, planning a product, or need independent proof, wait for the technical report and a third-party scorecard. Developers already using Claude Code can try it through its Anthropic-compatible local endpoint, though I haven't tested that pairing myself.
What this review means for support teams
The pitch that will reach support leaders is privacy: run the model on your own hardware and ticket data never leaves the building. I understand the pull. On sales calls, the data question is often the first one. A technical evaluator at a semiconductor-hardware company asked us whether the bot would fall back on "some other ChatGPT" when it didn't know an answer, and whether that could be turned off. A Danish telematics buyer worried about card numbers and passwords sitting inside tickets.
Neither concern was really about where the model runs. Both were about what the AI is allowed to read and say. A local model that answers from its general training is still the "other ChatGPT" that evaluator was worried about. What fixed it for them was an agent that answers only from approved knowledge and says when it doesn't know, plus redaction for sensitive fields.
That's how eesel works, and it's why I'd frame Edge0 as infrastructure while eesel is the employee. You hire an AI helpdesk teammate that joins your existing helpdesk, learns from your help center, macros, and past replies, and gets simulated against historical tickets before it talks to customers. Edge0's license would also block most of what a support vendor does with a model, since a commercial model service needs a separate agreement.
Two customers show what a careful setup looks like. Dokument 24, a Danish legal-documents service, has eesel draft Danish replies as internal notes from about 2,600 web pages and Zendesk items, and the note says so when the knowledge doesn't cover a question. smava, a German loan comparison company in regulated finance, runs about 27,000 Zendesk tickets a month through eesel, which categorised 14,600 and prepared 15,300 draft replies that wait for an advisor's approval.
For more on keeping bots grounded, see our guide to AI hallucinations in support, the MCP for customer support explainer, and the roundup of open-source chatbot platforms if self-hosting is a hard requirement.
Try eesel
If you're reading an Edge0-145B review because your support team wants AI without handing tickets to a black box, start with the part that actually decides privacy and accuracy: what the agent reads and when it hands off. eesel plugs into Zendesk, Freshdesk, and the other major helpdesks, answers only from your approved knowledge, and lets you watch it answer real past tickets before it goes live. Pricing is a fixed monthly credit plan, not per seat.
Developers can drive the same workspace from the eesel CLI. A coding agent, including one pointed at a local model, can upload approved knowledge such as exported help articles and saved replies, adjust the teammate's instructions, and ask test questions to check the answers. Connecting the helpdesk itself still needs someone with admin access to approve it.
Try eesel free, or book a demo to see it run on your own tickets.