The $0.197-per-task open-source agent model must first clear the VRAM hurdle

Bindu Reddy announced Smaug-Nemotron on X with a very direct framing: it is the strongest model in the agentic category among US open-source models, ranks first in agentic coding among US open-source models, is first in cost per task, has weights on Hugging Face, and can run locally on a desktop. The post also says it is built on Nvidia's latest model. Original post

The table is the most worth-looking part of this launch. Smaug-Nemotron-3.5-NVFP4 has AGENCY 19.6, DESIGN 62.7, CODING 72.2, BROWSER USE 8.0, REASONING 75.1, and a cost of $0.197 per successful task. Muse Spark Open has browser use 36.0 and a cost of $0.958. Beam 1.0 has coding 83.9 and reasoning 80.2, but browser use is 0.0 and cost is $1.662. Gemma 4 31B also costs $0.197, with agency 13.3 and coding 49.0. The table puts cheapness and capability in the same image, so readers can easily look at cost first and then at the weak spots.

The 8.0 in browser use stands out. X user AI Quanting asked directly whether this is a harness problem or whether the model just does not like driving a browser. This question is closer to everyday use than the claim of being first among US open-source models. When an agent clicks buttons, fills out forms, and handles login states in a browser, one failure can break the whole task chain. A high coding score does not mean tool calls and browser operations are equally stable. Another reply also mentioned that local models usually die on tool calls. The post does not elaborate on details beyond coding.

Desktop also needs to be looked at separately. Someone in the comments asked how many cards are needed, and Bindu Reddy replied, One DGX-Spark :). X user ＳＩＮＴ〔夜〕 did the math more concretely: NVFP4 weights are about 75GB and require a DGX Spark or H200, not a 4090. This information pulls desktop local running from an ordinary desktop PC back to workstation level. Christian Childers's reply was more direct: running local models is still impractical for ordinary users unless it is one-click install and does not eat 24GB of VRAM. A desktop can be a small AI workstation, or it can be the gaming card on everyone's desk; between the launch framing and users' understanding, there is VRAM in the gap.

The cost figure still has impact. At $0.197 per successful task, for teams that pay by API calls, this is a table that will be screenshotted over and over. X user Shinka said that a local model topping the cost-per-task ranking for agentic coding is a chart API pricing teams do not want to see. Hongji Lee added the value of open weights: when weights can run at home, no one can quietly nerf them next month. This angle is more practical than benchmark scores. Model capabilities will catch up; deployment rights and version stability are the long-term variables.

If you want to try it, check the hardware first. The post says the weights are on HF, and the NVFP4 quantized version is about 75GB. DGX Spark or H200 is the threshold given in the comments; ordinary 24GB graphics cards are not in the same discussion. Then check the task type. Agentic coding and reasoning are the strengths in this table, while browser use is a clearly low score. Running it through your own CRM, forms, and browser workflows is more useful than looking at the overall leaderboard. Gemma 4 31B is the same price and has much lower coding and agency, but its ecosystem and deployment experience may be more mature. The table only gives one slice.

Smaug-Nemotron is worth watching. It puts US open source, local weights, and cost per task into the same launch sentence, yet the comments quickly land on VRAM, browser use, and tool calls. This gap determines whether it can move from a leaderboard into everyday tasks. The launch framing can be loud; the answer from desktop fans is more honest.