# Is Smaug-Nemotron-3.5 SOTA?

Pending benchmarks. Announced Oct 23, 2026; no independent results yet.

As of Nov 4, 2026, Smaug-Nemotron-3.5 is pending benchmarks on isitsota: announced Oct 23, 2026, with no independent results on the boards we track yet.

Abacus.AI's agentic-coding fine-tune of NVIDIA's Nemotron 3.5 Super VL (120B total / 12B active), supervised-fine-tuned on multi-turn agentic coding trajectories. Open weights on Hugging Face in BF16 and NVFP4 under the OpenMDW-1.1 license inherited from the base; Abacus.AI says it runs on a single DGX Spark, H200 or GB200.

Canonical: https://isitsota.com/m/smaug-nemotron-3-5
Last updated: 2026-10-23
Tracking since: 2026-10-23

## Announced facts

| Fact | Value |
| --- | --- |
| Lab | Abacus.AI |
| Announced | Oct 23, 2026 |
| API | Not yet |
| Open weights | Published · Hugging Face: abacusai/Smaug-Nemotron-3.5 (BF16) and abacusai/Smaug-Nemotron-3.5-NVFP4; not gated |
| Parameters | 120B total / 12B active |
| Base model | nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 |
| Training | Supervised fine-tuning on agentic multi-turn coding trajectories |
| Architecture | Hybrid Mamba-2 + mixture-of-experts + attention, with a C-RADIO vision encoder and an MTP head (from the base) |
| Input / output | Text, images and video in; text out. Reasoning can be switched on or off |
| Context | Up to 1M tokens |
| License | OpenMDW-1.1 (inherited from the NVIDIA base) |
| Hardware | Lab says one DGX Spark, one H200 or one GB200 is enough |

Tags marked unsure are not stated on the lab's own pages.

## What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

- [x] Publicly announced with a source
- [x] Usable by the public (API or weights)
- [ ] Listed on an independent board we track
- [ ] Enough results for a rank in a category
- [ ] Holds across two snapshots (crowns only)

## What the lab claims

Self-reported. Shown for context; not used in any verdict.

- Lab claim (self-reported, model card, Smaug vs the base): SWE-bench Verified 72.3 vs 71.1; SWE-bench Multilingual 68.9 vs 67.7; Terminal-Bench 2.1 69.1 vs 60.7; DeepSWE 27.4 vs 24.7; NL2RepoBench 32.2 vs 30.2; LiveBench 2.0 Agentic 18.1 vs 15.8.
- Lab claim: the best US open-source agentic model that can run on a single machine (Abacus.AI CEO's announcement).

## Sources

- LAB: [Bindu Reddy (Abacus.AI CEO) announcing Smaug-Nemotron-3.5 on X](https://x.com/bindureddy/status/2113480704764076310) (2026-10-23)
- REPO: [abacusai/Smaug-Nemotron-3.5 on Hugging Face](https://huggingface.co/abacusai/Smaug-Nemotron-3.5) (2026-10-23)
- REPO: [abacusai/Smaug-Nemotron-3.5-NVFP4 on Hugging Face](https://huggingface.co/abacusai/Smaug-Nemotron-3.5-NVFP4) (2026-10-22)

Tracking continues until an independent board lists it.

## Related

- Lab: [Abacus.AI](https://isitsota.com/labs/abacus-ai.md)
- All pending releases: https://isitsota.com/releases.md
- JSON: https://api.isitsota.com/v1/models/smaug-nemotron-3-5