Edge0-145B needs 29GB peak memory, but its licence is unclear and its benchmarks are self-reported

The release also includes a smaller Edge0-120B, a unified inference framework, deployment tools and the Edge0 App. Edge0 calls Edge0-145B the world's first open-source on-device model with 145 billion parameters. On Hacker News, one commenter asked whether the release is derived from Qwen3.8-Flash-Next. A reply noted that the Hugging Face page lists Qwen3.8-Flash-Next as the base model. Another commenter described the inference engine, edge0-strata, as a fork of strata.

One commenter on Hacker News worked out the size of the weights. At 4 bits, 145B parameters take 72.5GB, which left the commenter unsure the whole model fits in 29GB. Edge0's post describes loading across three storage tiers instead. Frequently used hot experts stay in GPU memory. An expert library of approximately 40GB sits in system memory and is brought into GPU memory as needed, and approximately 48GB of vocabulary data lives on the SSD and is read on demand.

The post does not say which memory the 29GB peak measures, or how a 40GB expert library in system memory fits within it. It is also inconsistent about the peak itself: the laptop comparison elsewhere in the post gives 28GB. Until that is cleared up, size a machine for the stated peak plus the expert library in system memory, plus roughly 48GB of free SSD. Edge0 says its technical report will include the evaluation setup, hardware environment and reproduction instructions.

The main architectural change is in the sparse experts. Each group of 64 experts shares two large matrices, and each expert keeps only a small set of parameters that captures how it differs. Previously, every expert stored a complete set of matrices. For scheduling, the GPU computes the cached experts while the CPU handles the remaining work in parallel. Speculative decoding verifies several candidate tokens at once, which spreads the cost of reading weights across more output tokens, and Edge0 says these choices support the 80+ tokens/s decode rate.

Edge0 says it trains the model in its on-device architecture from the outset, rather than compressing or pruning a cloud model after training. That claim sits uneasily with the base-model listing on Hugging Face that one commenter pointed to. Another commenter's reading is that Edge0 fine-tuned Qwen3.8-Flash-Next, made some tweaks to the model and forked strata to run it. The sources do not settle which account is correct. The answer affects how the gains over Qwen should be read and which licence terms apply.

Commenters on Hacker News linked an NVFP4 build and a q3k build of the weights on Hugging Face. One commenter said NVFP4 restricts that build to Nvidia Blackwell, and presumed the inference engine requires CUDA 13. The same commenter noted that no standard GGUF files have been published. The variants link instead leads to a directory with the model split across many files.

The commenter who quoted that line said there is no guidance on how to run the model on a Mac. A second commenter said Edge0 benchmarked an M2 Ultra but is not releasing a Mac runtime today. The sources do not resolve the conflict between the stated Apple M-series support and these reports. Treat the MacBook Pro in Edge0's memory table as an example of a compatible laptop, not as a confirmed download path, until a Mac runtime ships.

Edge0 describes the release as open-sourcing the models and releasing the full stack, but the available text of the blog post does not name a licence. One commenter on Hacker News says the licence is too restrictive to meet the standard definition of open weight. Another quoted the model page, which says use may be subject to Qwen's separate authorization or attribution requirements. Teams planning to ship a product on these weights should read the model card's licence text, and any Qwen terms it inherits, before writing integration code.

Qwen3.8-Flash-Next is the closest model to Edge0-145B by six-benchmark mean. Against it, Edge0-145B leads on DeepSWE 1.1 (64.6 vs 58.7), Terminal-Bench 2.1 (86.52 vs 83.3) and, narrowly, AA-LCR (79.3 vs 79.1). It trails on HLE (35.4 vs 35.9), GPQA-Diamond (90.4 vs 91.7) and IFBench (80.6 vs 81.3). DeepSeek-V4-Flash-0731 scores 54.4 on DeepSWE 1.1, 82.7 on Terminal-Bench 2.1, 90.8 on GPQA-Diamond and 79.2 on IFBench.

Edge0's prose compares its GPQA-Diamond and IFBench results only against DeepSeek-V4-Flash-0731 and Claude-Opus-4.6 (Max). Qwen3.8-Flash-Next, which scores higher on both, is left out of those comparisons. Claude-Opus-4.6 (Max) is ahead on HLE at 40.0 and on GPQA-Diamond at 91.3, but behind on Terminal-Bench 2.1 at 78.2 and IFBench at 62.5. Outside the Edge0-evaluated rows, Flint 40B posts 89.2 on IFBench and Muse Glimmer posts the table's highest AA-LCR score at 83.3.

Not all of the numbers come from the same harness. The Flint, Muse Glimmer and Gemma 4 rows mix Terminal-Bench 2.1 and IFBench results from the Recursive Flint evaluation with HLE, GPQA and AA-LCR results from Artificial Analysis. Claude-Opus-4.6 has no public DeepSWE 1.1 score. Against Qwen3.8-27B, Edge0's relative gains are 53.1% on DeepSWE and 18.5% on Terminal-Bench, against 14.9% on HLE, 1.4% on IFBench and 1.3% on GPQA. The DeepSWE lead over Qwen3.8-Flash-Next is 5.9 points, also from internal evaluation.

One commenter argued that Terminal-Bench scores often track real-world performance for coding agents that rely heavily on bash, which makes the 86.52 worth testing. Another said the release is exciting if third-party testing holds up the published values, but was skeptical given the model size. No third-party results appear in the sources. The technical report and its reproduction instructions have not yet been published.

Edge0's headline efficiency figure is a 72.8 mean at 29GB. Qwen3.8-Flash-Next averages 71.7 at approximately 70GB, and DeepSeek-V4-Flash-0731 averages 70.1 at approximately 87GB. Edge0 says this puts it at about 40% of the estimated baseline memory. The baseline footprints are estimates for a standard 4-bit deployment that the footnote says remain to be measured, while the Edge0 figures are measured internal peaks.

Edge0-120B averages 68.9 at the same 29GB, and Qwen3.8-27B averages 66.1 at an estimated 18GB. One commenter suggested that if the Qwen comparison uses a quantized version, part of Edge0's lead over its apparent base model could come from Edge0's format losing less to quantization. The sources contain no measured memory figure for any baseline.