The Pulse
Abacus Releases a 4-Bit Nemotron Fine-Tune for Coding Agents
Abacus.AI released a four-bit agentic coding fine-tune of NVIDIA’s Nemotron 3.5 Super VL, reporting gains over the NVFP4 base on agentic coding benchmarks. The company published the weights on Hugging Face, while an October 23 tracker liste
AI.info Team · Written with AI · fact-checked ·
Abacus brings its Nemotron fine-tune down to four bits
Abacus.AI published Smaug-Nemotron-3.5-NVFP4 on October 22, a four-bit version of its agentic coding fine-tune of NVIDIA’s Nemotron 3.5 Super VL. The base model has 120 billion parameters, with 12 billion active at a time; it combines Mamba-2, mixture-of-experts layers and attention, and accepts text, images and video. Abacus says the new checkpoint occupies about 75 GB, roughly a third of the BF16 model’s size.
That smaller footprint changes the hardware needed to serve the model. Abacus’s post describes NVIDIA’s corresponding NVFP4 checkpoint as running on one DGX Spark, one H200 or one GB200. It also says existing stacks serving that checkpoint can use Smaug as a drop-in replacement. The weights are published on Hugging Face under the OpenMDW 1.1 license inherited from NVIDIA’s base model.
Agentic scores rise against the four-bit base
Abacus reports improvements over the NVIDIA NVFP4 checkpoint on all five agentic coding tests listed in its model card. Smaug scores 71.8 on SWE-bench Verified, compared with 71.1 for the base, and 64.0 on Terminal-Bench 2.1, up from 62.8. Its DeepSWE score is 21.2 against 18.2, while SWE-bench Multilingual rises from 67.0 to 69.0 and NL2RepoBench from 28.3 to 30.2.
On LiveBench 2.0, the company reports 72.2 for Coding versus 61.3 for the NVFP4 base, and 19.6 for Agentic versus 11.5. The overall score is 52.9, up from 50.0. The model card says LiveBench scores for all three models are from the public leaderboard, while Smaug’s other figures are company runs; it also describes the evaluation setups and run counts.
The reported gains are concentrated in agentic coding rather than general capabilities. Abacus’s comparisons put reasoning, coding, instruction-following, long-context and vision results near the base model’s scores. The results are not an independent confirmation: an October 23, 2026, benchmark tracker lists no independent results for the model yet.
Training against the weights that will ship
Abacus says ordinary fine-tuning followed by four-bit quantization would erase much of the small update it wanted to preserve. Its approach, which it calls merge-aware training, runs each training step using weights rounded into the format intended for release. The learning objective rewards tool-call actions while applying a separate penalty to keep the model’s reasoning close to the base.
“The new idea here is that the loss always sees the model exactly as it will be served.”
Sandeep Singh Sandha and Arvind Sundararajan, the authors, write
The method adapts the language model, according to the company; the vision encoder, vision-language connector, MTP head and quantization layout remain those of NVIDIA’s checkpoint. That leaves the fine-tune’s main change in how the model behaves during multi-turn tool use, not in its architecture.
Open weights, with a self-reported scorecard
The model card recommends temperature 1.0 and top-p 0.95 for use. For coding agents, Abacus also advises retaining prior reasoning in the conversation history, using the setting truncate_history_thinking: false. The model supports vLLM, SGLang and TensorRT-LLM, according to the company’s release post.
Smaug-Nemotron-3.5-NVFP4 extends Abacus’s earlier BF16 fine-tune of the same NVIDIA model, rather than replacing it. Developers can compare the two published checkpoints, but the benchmark figures remain Abacus’s reported results until independent evaluations appear.