Edge0-145B / Edge0-120B

Edge0-145B and Edge0-120B are the on-device models Edge0 officially open-sourced on 12 November 2026. In the official blog Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer, dated November 12, 2026, the company wrote: “Today, we're open-sourcing Edge0-145B and Edge0-120B, together with our unified inference framework and deployment tools. Edge0-145B is the world's first open-source on-device AI model with 145 billion parameters.” Officially, they are 100B-class open-weight models meant to run on a personal computer for local coding, reasoning, and agent work, released together with the inference framework and deployment tools.

At 18:00:13 UTC the same day, @SamuelZengML posted “Introducing Edge0-145B and Edge0-120B,” calling them @Edge0_AI’s first device-native, open-weight 100B-class foundation models. At 18:01:27 UTC, the official @Edge0_AI account posted: “Frontier-level agentic coding, now running fully on-device. Edge0-145B and Edge0-120B are available with open weights.” The same account pointed to the official blog at 18:24:01 UTC. The weights appeared on Hugging Face earlier: Edge0/Edge0-145B-q3k has createdAt 2026-11-11T05:23:10Z and lastModified 2026-11-12T13:14:54Z. The public Introducing blog and official-account posts clustered around 18:00 UTC on 12 November (02:00 Beijing time on 13 November). The 04:32 and 08:32 scans that day did not store this official launch; this pass adds it.

According to company test figures in the official blog, Edge0-145B reports about 29GB peak memory on a compatible laptop, more than 1,100 tokens/s prefill, more than 80 tokens/s decode, typically around 120W, and a 128K–256K context; of the 145B parameters, about 10B participate in each step. Internal scores listed include 64.6 on DeepSWE 1.1 (versus 58.7 for Qwen3.8-Flash-Next and 54.4 for DeepSeek-V4-Flash-0731), 86.52 on Terminal-Bench 2.1 (versus 78.2 for Claude-Opus-4.6 Max), 90.4 on GPQA-Diamond, 80.6 on IFBench, and 79.3 on AA-LCR; Edge0-120B is listed in the same table at 53.6, 81.3, 87.4, 79.3, and 81.0. Officially, the models are trained in the on-device architecture from the start, share large matrices across sparse experts, and load across GPU memory, system memory, and SSD. The Edge0-AI/Edge0 repository describes the 145B / 120B line as latent MoE models derived from Qwen3.8-Flash-Next, trained natively at 4 bits, with Q3_K and NVFP4 weights, running through edge0-infer serve / chat on Linux or Windows with an NVIDIA GPU and CUDA 13. Hugging Face pages include Edge0-145B-q3k, Edge0-145B-NVFP4, and the matching 120B repos. Officially, this laptop tier follows the September 2026 Edge0-35B phone release.

Key points:

• Official blog: https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer ; official post: https://x.com/Edge0_AI/status/2120940049391100394 (2026-11-12T18:01:27Z); introducing post: https://x.com/SamuelZengML/status/2120939739431972922 ;
• Weights and code: https://huggingface.co/Edge0/Edge0-145B-q3k ; https://github.com/Edge0-AI/Edge0 ;
• These are concrete Edge0 model names. Do not confuse them with the earlier Edge0-35B / Edge0-8B models, or with the Audio8 family discussed previously. Speed, memory, and benchmark figures in the blog are company tests; this page does not independently reproduce them.

Why it matters

This is a new model-name pair Edge0 announced on 12 November 2026 through an official blog and official account: 100B-class open-weight on-device models meant to run on a personal computer, released with weights and an inference framework. The weight repos were created about a day before the public blog, but the Introducing post and the @Edge0_AI announcement landed at 18:00 UTC on 12 November; the 04:32 and 08:32 scans did not store them, and this pass adds the official launch. Do not treat this as merely a quantized pack of the existing Edge0-35B.

This page only restates official public material. The 29GB memory figure, 80+ tokens/s, DeepSWE / Terminal-Bench scores, and the claim that Edge0-145B is the first 145B open-source on-device model are company statements. This page does not independently reproduce or rank them.

Limits

The published score table comes from Edge0’s internal evaluations, including the comparison scores it lists, and the company says full reproduction details will accompany a technical report. The GitHub README currently points the 145B / 120B line at NVIDIA GPUs (CUDA 13) on Linux or Windows; the blog also mentions day-one support for macOS and Apple M-series chips. Follow the official repo and hardware guides rather than assuming every laptop can run it. Do not treat this entry as a guarantee of consumer-laptop performance, benchmark scores, or license terms. Do not describe these models as the same SKU as Edge0-35B or Audio8.

1. Read the official blog at https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer , and cross-check the official post at https://x.com/Edge0_AI/status/2120940049391100394 , the introducing post at https://x.com/SamuelZengML/status/2120939739431972922 , the Hugging Face weights at https://huggingface.co/Edge0/Edge0-145B-q3k , and the GitHub repository at https://github.com/Edge0-AI/Edge0 ;
2. Follow the official repository instructions to load the Q3_K or NVFP4 weights with edge0-infer or the Edge0 App on supported hardware, and read the official hardware and deployment guides first;
3. When citing the 29GB memory figure, 80+ tokens/s, DeepSWE, Terminal-Bench, or other comparison scores, attribute them to the company tests in the 12 November 2026 official blog and recheck the upstream source;
4. Do not describe Edge0-145B / Edge0-120B as Edge0-35B, Edge0-8B, or Audio8; this page only restates the official blog and announcement posts — recheck the upstream sources before citing them externally.