Abacus.AI has released Smaug-Nemotron-3.5, an open-weight model built by fine-tuning Nvidia’s Nemotron 3.5 Super for agent work, and on Abacus’s own LiveBench 2.0 leaderboard its 4-bit version costs $0.197 per successful task. Chief executive Bindu Reddy announced it on X on 24 October 2026, calling it “the best US open-source model in the agentic category”. The model card on Hugging Face, created on 18 October 2026, gives the details.
The same leaderboard puts the model’s agentic score at 19.6 out of 100, against 50.9 for DeepSeek V4.1 Flash 1017. LiveBench 2.0 is run by Abacus: its page says “We use these results to power our powerful agent and supercomputer, available via an Abacus AI subscription.”
Business Pill · AI AGENTS
A short Business Engineer explainer on AI agents that do work rather than answer questions. It teaches the general idea only and says nothing about the model in this piece.
The key insight: As we read it, the number to repeat is the gap on Abacus’s own board: its new model scores 19.6 on agentic work against 50.9 for DeepSeek V4.1 Flash 1017, at almost the same cost per successful task, $0.197 against $0.202.
The base. The model card describes it as an agentic fine-tune of Nvidia’s Nemotron 3.5 Super VL 120B-A12B, which it calls “a 120B-parameter (12B active) hybrid Mamba-2 / MoE / Attention vision-language model”. It keeps the base’s context of up to 1 million tokens and its licence, the OpenMDW License Agreement, version 1.1.
What changed. “The fine-tune adapts behaviour only,” the card says, and “Every architectural parameter is unchanged from the base”. Training data “consists of filtered multi-turn, tool-using coding trajectories.”
The gains Abacus reports. On Terminal-Bench 2.1 the card lists 69.1 against 60.7 for the base, and on SWE-bench Verified 72.3 against 71.1. The card says the Smaug scores “are our runs”, while the base scores are as reported on Nvidia’s own card.
The hardware. A second card, for a 4-bit version created on 22 October, says “4-bit weights put the full 120B model in about 75 GB, roughly a third of its BF16 size, so it fits on a single H200, two H100 or one GB200.”
Agentic score out of 100 on LiveBench 2.0, a leaderboard run by Abacus.AI, for selected open models: DeepSeek V4.1 Flash 1017, GLM-5.3 Flash, the 4-bit Smaug-Nemotron-3.5, Muse Spark Open and the base Nemotron 3.5 Super VL. Source: livebench.ai, last updated 23 Oct 2026, read 24 Oct 2026.
What Was Posted on X
๐จ Smaug-Nemotron is the best US open-source model in the agentic category
– can run locally on your desktop! – tops agentic coding in US open-source – number #1 on cost/task – builds on Nvidia’s latest model
We are excited to push the boundaries of US open-source. Weights available on HF
Bindu Reddy, chief executive of Abacus.AI, announces Smaug-Nemotron in this post, calling it the best US open-source model in the agentic category, saying it can run locally on a desktop, is number one on cost per task and builds on Nvidia’s latest model, with weights on Hugging Face. These are her claims; the piece sets them against Abacus’s own model card and LiveBench 2.0 board.
The Model Card’s Numbers
Scores out of 100, from the Smaug-Nemotron-3.5 model card. Smaug scores are Abacus’s runs; base scores are as reported on Nvidia’s card, except the LiveBench 2.0 row, which the card says is the official LiveBench 2.0 leaderboard entry.
Benchmark
Smaug-Nemotron-3.5
Nemotron 3.5 Super VL (base)
Terminal-Bench 2.1
69.1
60.7
SWE-bench Verified
72.3
71.1
SWE-bench Multilingual
68.9
67.7
DeepSWE (offline)
27.4
24.7
LiveBench 2.0, Agentic
18.1
15.8
GPQA Diamond
90.7
89.5
Source: Hugging Face model card, abacusai/Smaug-Nemotron-3.5, read 24 October 2026.
On Abacus’s Own Leaderboard
LiveBench 2.0, last updated on 23 October 2026, lists 53 models. Its page says “We give models real work (decks, research reports, bug fixes, apps that run for hours) and grade what they hand back.” The 4-bit Smaug-Nemotron-3.5 has an overall score of 52.9, the 45th row of 53 when the table is sorted by overall score. Its agentic score is 19.6 and its cost per successful task is $0.197.
The chart Reddy posted shows seven open models from that board, with Smaug first on the agentic column at 19.6, ahead of Muse Spark Open at 18.9. On the full board, DeepSeek V4.1 Flash 1017 scores 50.9 on agentic at $0.202 per successful task, and GLM-5.3 Flash scores 28.4 at $0.153. The highest overall score is Claude Fable 5.1 at 88.0, at $5.096 per successful task.
Reddy’s post calls the model “number #1 on cost/task”. In the seven-model chart, $0.197 is the lowest figure, shared with Gemma 4 31B. On the full board, several models list a lower cost per successful task, including GPT-6 Luna at $0.127 and GLM-5.3 Flash at $0.153.
By Our Arithmetic
By our arithmetic, on LiveBench 2.0 the 4-bit Smaug model scores 3.8 points above the base Nemotron 3.5 Super VL on agentic (19.6 against 15.8), at a cost per successful task about 34% lower ($0.197 against $0.300). DeepSeek V4.1 Flash 1017’s agentic score of 50.9 is about 2.6 times Smaug’s 19.6, at a cost per successful task within a cent of it.
Connecting the Coverage
The release lands in an open-model race led from China. Our Kimi piece reported Moonshot pricing its open-weight Kimi K3.1 Pro at $2.00 per million input tokens and $10.00 per million output tokens, half of Claude Opus 5.5. Our China datacentre piece added that OpenRouter’s rankings, read on 24 October, list the five most-used models by tokens that week as all from Chinese developers, with Claude Opus 5.5 sixth.
Our Nvidia open-models piece reported Nvidia saying on 21 October that SAP, Salesforce and others are working with its open Nemotron models “to create custom models and build domain-specific agents.” And our Meta tool-selection piece covered a paper in which a trained 8B model picked the right tools for 41.5% of test requests, against 30.6% for a prompted 402-billion-parameter Llama model.
What’s new: Abacus has published an agent-tuned version of Nvidia’s open Nemotron model and ranked it on a leaderboard it runs itself. There it leads the seven open models Reddy charted and, by our arithmetic, trails the top DeepSeek model by 31.3 points on agentic work.
The Structural Read
As we read it, the release shows where Nvidia’s open Nemotron models fit: as a base that other companies fine-tune and publish. The card says the fine-tune “adapts behaviour only” and that any stack serving the base “serves this model as a drop-in replacement.”
As we read it, the measurement and the product come from the same company: LiveBench 2.0 is run by Abacus, and its page says the results power Abacus’s own agent, sold by subscription. The piece’s figures are therefore Abacus’s, not an outside test.
As we read it, the cost column is the business detail: on the board, the 4-bit Smaug model lists $0.197 per successful task against $0.300 for the base, while frontier models such as Claude Fable 5.1 list $5.096.
Bindu Reddy, chief executive of Abacus.AI, on X, 24 October 2026
“We are excited to push the boundaries of US open-source. Weights available on HF”
Three Implications
FOR TEAMS BUILDING AGENTS The 4-bit card says the 120B model fits on a single H200, two H100 or one GB200, and runs on any stack that serves the Nvidia base.
FOR NVIDIA Its open Nemotron base now has a published agent fine-tune from a US company, after Nvidia said SAP and Salesforce are working with Nemotron models.
FOR BUYERS COMPARING OPEN MODELS On the same board, the top DeepSeek model scores 50.9 on agentic work at $0.202 per successful task; the Smaug model scores 19.6 at $0.197.
The framework states: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, the model weights here are the open layer: Nvidia publishes the Nemotron base and Abacus publishes its fine-tune. What Abacus sells, per its leaderboard page, is the agent built on top, by subscription. The framework was published in March 2026 and does not discuss this release.
What Is Not Established
All Smaug scores come from Abacus, either from its own runs on the model card or from the LiveBench 2.0 board it operates; we found no independent evaluation in what we read. Reddy’s post says the model “can run locally on your desktop”; the model card, in what we read, names a single H200, two H100 or one GB200, which are data-centre GPUs. The board does not list a country for each model, and the seven-model chart’s selection rule is not stated on it. We did not contact Abacus or Nvidia.
Business Engineer Framework
The Open vs Closed Meta-Framework
A Business Engineer framework on which layers of a technology stack to open and which to keep closed.
Abacus.AI’s Smaug-Nemotron-3.5, a fine-tune of Nvidia’s 120-billion-parameter Nemotron 3.5 Super, scores 69.1 against the base’s 60.7 on Terminal-Bench 2.1 in Abacus’s runs. On Abacus’s LiveBench 2.0, its 4-bit version scores 19.6 on agentic work at $0.197 per successful task, first among the seven open models in Reddy’s chart and 45th of 53 on overall score.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This piece rests on Abacus.AI’s two model cards on Hugging Face, the LiveBench 2.0 leaderboard page (run by Abacus.AI, last updated 23 October 2026), and Bindu Reddy’s post on X, all read on 24 October 2026, plus our own earlier pieces. Benchmark scores are Abacus’s own runs or leaderboard entries and were not independently reproduced. We did not contact Abacus or Nvidia.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
