On November 12, on-device model company Edge0 released open weights for Edge0-145B and Edge0-120B, along with its C++ inference engine edge0-infer and Edge0 Chat 0.2.0. The 145B model is built on Qwen3.8-Flash-Next as a latent sparse MoE, with about 10.12B active parameters per token, native 4-bit training and a 256K context. Edge0 reports 29GB peak memory locally and 98.3 tok/s agent decode on an RTX 5090 D 24GB. It reports 86.52 on Terminal-Bench 2.1, versus 78.2 for Claude Opus 4.6 (Max). All scores come from Edge0 internal evaluations.
Key Takeaways
- ✓Size and structure: Edge0-145B has 145B total parameters and about 10.12B active per token; 48 layers with 512 experts each, 10 routed plus 1 shared active; 256K context, trained natively at 4 bits
- ✓Benchmarks (vendor internal evals): 86.52 on Terminal-Bench 2.1, versus 83.3 for Qwen3.8-Flash-Next, 82.7 for DeepSeek-V4-Flash-0731 and 78.2 for Claude Opus 4.6 (Max); 64.6 on DeepSWE 1.1 and 90.4 on GPQA-Diamond
- ✓Local speed (model card figures): the NVFP4 build reaches 98.3 tok/s agent decode and 3165.7 tok/s prefill on a 16K input on an RTX 5090 D 24GB; the Q3_K build hits 91.7 tok/s agent decode on an RTX 4090
- ✓Hardware floor: the Windows package needs an NVIDIA RTX 20-series or newer GPU with at least 16GiB of VRAM plus 20GiB of system RAM; weights are 47.79GiB for NVFP4 and 40.81GiB for Q3_K
- ✓License limits: the model is under the Edge0 Community Model License 1.0 plus the Qwen Community License; personal, research and internal use are allowed, but reselling the model or running a commercial model service needs a separate license; the inference framework is Apache-2.0
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the problem
On-device models have long been limited by memory. Edge0's blog notes that before September 2026 most on-device models had under 10B parameters: Apple's on-device foundation model had about 3B, and Gemini Nano came in 1.8B and 3.25B versions. Running a local agent that writes code and calls tools meant either a small model or paying cloud token fees. Edge0, founded by Samuel Zeng, shipped the phone-scale Edge0-35B in September. This release pushes to 145B, aiming to run a near-cloud-level coding agent on a laptop or a single gaming GPU.
Architecture and how it works
Edge0-145B is built on Qwen3.8-Flash-Next as a latent sparse MoE. It has 48 layers with 512 experts each; every token activates 10 routed experts plus 1 shared expert, about 10.12B parameters in total. Each group of 64 experts shares two large matrices while each expert keeps a small set of its own parameters, for an 84.6B expert pool. A 51.2B n-gram embedding table is stored in fp8 on the SSD and read on demand, and a 1.26B draft head handles speculative decoding. At runtime hot experts stay in GPU memory, an expert library of about 40GB sits in system RAM, and the GPU and CPU compute in parallel. The model is trained at 4 bits from the start rather than quantized afterward.
Benchmarks and measured results
All figures are vendor-reported: Edge0, Qwen, DeepSeek and Claude scores come from Edge0's internal evaluations, and the technical report is not out yet. Terminal-Bench 2.1: Edge0-145B 86.52, Qwen3.8-Flash-Next 83.3, DeepSeek-V4-Flash-0731 82.7, Claude Opus 4.6 (Max) 78.2. DeepSWE 1.1: 64.6 versus 58.7 and 54.4. GPQA-Diamond is 90.4, slightly below Claude's 91.3; HLE is 35.4, below Claude's 40.0. The six-benchmark mean is 72.8, versus 71.7 for Qwen3.8-Flash-Next. On speed, the model card reports 98.3 tok/s agent decode for the NVFP4 build on an RTX 5090 D 24GB, and 60.9 tok/s chat decode for the Q3_K build on an Apple M2 Ultra.
Getting started for developers
Weights are under the Edge0 organization on Hugging Face: 47.79GiB for NVFP4 and 40.81GiB for Q3_K. Edge0 Chat 0.2.0 currently ships only for Windows and needs an NVIDIA RTX 20-series or newer GPU with at least 16GiB of VRAM, 20GiB of system RAM and CUDA Toolkit 13. edge0-infer provides run, chat, bench and serve commands, and its local server speaks both the OpenAI API and Anthropic's /v1/messages, so it can plug into existing coding agents. Read the license before commercial use: the model cannot be resold or used for a commercial model service or a standalone commercial coding assistant without a separate Edge0 license. The vendor scores still await third-party reproduction.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.