The most important thing about Edge0-145B is often buried under the headline: it is not a from-scratch base model but a derivative of Alibaba's Qwen3.8-Flash-Next — the model card's metadata names base_model as Qwen/Qwen3.8-Flash-Next, with the routed experts in all 48 layers rebuilt as latent MoE and post-training afterwards on Edge0's own data. The right question is not "how strong is it" but "what did the rebuild and post-training buy over the base". Every score below is vendor-reported (Edge0 internal evaluations) with no independent verification as of Nov 20, 2026.
| Benchmark | Edge0-145B | Qwen3.8-Flash-Next | Delta |
|---|---|---|---|
| DeepSWE 1.1 | 64.6 | 58.7 | +5.9 |
| Terminal-Bench 2.1 | 86.52 | 83.3 | +3.22 |
| HLE | 35.4 | 35.9 | −0.5 |
| GPQA-Diamond | 90.4 | 91.7 | −1.3 |
| IFBench | 80.6 | 81.3 | −0.7 |
| AA-LCR | 79.3 | 79.1 | +0.2 |
| Mean (6) | 72.8 | 71.7 | +1.1 |
What the rebuild bought is direction, not breadth. DeepSWE and Terminal-Bench are agent and terminal-coding tasks — exactly Edge0-145B's positioning — and together they contribute +9.1 points, the whole of the rebuild's gain. HLE, GPQA-Diamond and IFBench — knowledge and instruction-following — all trail, by −2.5 points combined, which suggests the post-training data mix pushed capability toward coding while giving back a little general ability. AA-LCR is essentially level. The +1.1 mean is the net of those two forces.
There is also a structural difference the scores do not capture: the base Qwen3.8-Flash-Next is a cloud model, while Edge0-145B was rewritten for on-device use — natively trained at 4 bits, latent MoE, a three-tier memory pipeline, 256K context. The score lead is only 1.1 points, but the deployment shape is entirely different: one needs a data centre, the other one consumer GPU plus a large helping of RAM. The reason to buy Edge0-145B is first "a 145B that runs locally" — the score lead is secondary.
| Benchmark | Edge0-145B | DeepSeek-V4-Flash-0731 | Delta |
|---|---|---|---|
| DeepSWE 1.1 | 64.6 | 54.4 | +10.2 |
| Terminal-Bench 2.1 | 86.52 | 82.7 | +3.82 |
| HLE | 35.4 | 33.8 | +1.6 |
| GPQA-Diamond | 90.4 | 90.8 | −0.4 |
| IFBench | 80.6 | 79.2 | +1.4 |
| AA-LCR | 79.3 | 79.7 | −0.4 |
| Mean (6) | 72.8 | 70.1 | +2.7 |
The mean lead over DeepSeek-V4-Flash is +2.7, larger than the +1.1 over the base — but the shape is the same: the lead concentrates on DeepSWE (+10.2) and Terminal-Bench (+3.82), while GPQA-Diamond and AA-LCR still trail. In other words, Edge0-145B is "strong on coding, level on knowledge" against everyone — not an artefact of the comparison model, but the shape of its own capability. Note that DeepSeek-V4-Flash-0731 is a July 31 build, a version more than three months old.
Edge0-120B launched the same day as the 145B — the smaller tier of the same line. But as of Nov 20, 2026, neither the official blog nor the model cards publish the 120B's full specs or a six-benchmark comparison table. What this site's research confirms is only that it launched the same day, is likewise a latent-MoE derivative of Qwen3.8-Flash-Next, and likewise ships in Q3_K and NVFP4 builds. With no official 120B benchmark numbers, this site makes no quantified claim about "how much stronger the 145B is than the 120B"; which tier to pick awaits official 120B specs and scores, or independent testing. One warning: some third-party pieces quote Terminal-Bench figures for the 120B, but the numbers contradict each other across sources (75.3 vs 81.3) — none of them is used until officially confirmed.
Local agent and coding users: pick Edge0-145B. Its two leads (DeepSWE, Terminal-Bench) land exactly on your workload, and it is the only one of the three that runs locally — the base and DeepSeek-V4-Flash have no equivalent on-device form.
Knowledge Q&A and instruction-following first: the base may not be worse. The base leads on all three of HLE, GPQA-Diamond and IFBench; if your work runs in the cloud with no local-deployment need, there is no reason to switch models for a +1.1 mean.
On a budget and curious about on-device 145B: look at the 120B first. Until official 120B data is published, the value-for-money calculation cannot be done; decide big or small after the official figures or independent tests arrive.
Reference: Edge0 — Official blog "Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer" · Hugging Face — Edge0/Edge0-145B-q3k model card · Hugging Face — Edge0/Edge0-145B-NVFP4 model card · GitHub — Edge0-AI/Edge0 (rewritten and organized by ModelBrief; facts per official sources)