1. Model Introduction
Smaug-Nemotron-3.5 is an agentic fine-tune of NVIDIA Nemotron 3.5 Super VL 120B-A12B, a 120B-parameter (12B active) hybrid Mamba-2 / MoE / Attention vision-language model, fine-tuned by Abacus.AI. It improves on the base on agentic coding (+8.4 on Terminal-Bench 2.1, +2.7 on DeepSWE, +2.3 and +0.7 on the LiveBench 2.0 Agentic and Coding boards, +1.2 on SWE-bench Verified and +1.2 on SWE-bench Multilingual) and holds reasoning, coding, instruction-following, long-context and vision scores at parity with the base (GPQA Diamond, LiveCodeBench v6, SciCode, IFBench, AA-LCR and V*).
The fine-tune adapts behaviour only. Every architectural parameter is unchanged from the base: same 120B hybrid MoE, same vision encoder, same tokenizer, same chat template, same reasoning on/off switch. Any inference stack that serves Nemotron 3.5 Super VL BF16 serves this model as a drop-in replacement.
This card describes the training approach and the evaluation results. Training data consists of filtered multi-turn, tool-using coding trajectories.
2. Model Summary
| Base Model | nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 |
| Total Parameters | 120B (12B active) |
| Architecture | Mamba-2 + MoE + Attention hybrid with a C-RADIO vision encoder and an MTP head (unchanged) |
| Inputs / Outputs | Text, images and video / text (as the base) |
| Precision | BF16 |
| Context Length | Up to 1M tokens (as the base) |
| Reasoning Mode | On by default; configurable via chat template (enable_thinking=True/False) |
| Recommended Sampling | Temperature 1.0, Top_P 0.95 (vision: temperature 0.6, as the base) |
| Adaptation | Supervised fine-tuning on agentic trajectories |
| License | OpenMDW License Agreement, version 1.1 (inherited) |
3. Evaluation
Smaug-Agentic-Nemotron vs its base
Higher is better. Base and Nemotron 3 Ultra scores from the Nemotron 3.5 Super VL model card and livebench.ai.
| Smaug-Nemotron-3.5 | Nemotron 3.5 Super VL |
|---|
LiveBench 2.0 category profile
Scores 0–100 on LiveBench 2.0; overall = mean of the seven boards. Base scores are the published leaderboard entry.
NotesAll results were produced with vLLM serving the BF16 weights (FP8 KV cache, no MTP speculative decoding) on NVIDIA B300 GPUs, temperature = 1.0, top_p = 0.95 (AA-LCR: top_p 1.0, as in NVIDIA's recipe; V*: temperature 0.6, the base's vision setting), thinking enabled, following the NVIDIA Nemotron 3.5 Super reproducibility recipes wherever one exists.
- SWE-bench Verified: NVIDIA NeMo Gym recipe (OpenCode agent), average of 3 runs over all 500 tasks.
- SWE-bench Multilingual: NVIDIA NeMo Gym recipe (OpenCode agent), average of 6 runs over all 300 tasks, scored with NVIDIA's grader as is.
- Terminal-Bench 2.1: NVIDIA NeMo Gym recipe with the Terminus 2 agent, average of 4 runs over all 89 tasks, with past reasoning kept in the conversation history (see Section 7).
- DeepSWE [offline]: NVIDIA NeMo Gym DeepSWE environment with the OpenCode agent, the harness behind the base card's "DeepSWE [offline]" row; pass@1 over all 113 tasks, single run.
- NL2RepoBench: the benchmark's official OpenHands 0.56 scaffold (NVIDIA publishes no recipe), mean fraction of hidden pytest tests passed over all 104 tasks, up to 500 agent steps with 196k input and 64k output tokens per step, single run; the base was run by us under the identical stack, and the difference is within run-to-run variance.
- LiveCodeBench v6: official recipe, 454 problems, pass@1, average of 4 runs.
- SciCode: official recipe, subtask accuracy over 288 sub-steps (65 problems), average of 3 runs.
- AA-LCR v1.1: official recipe, 100 questions, average of 5 runs.
- V*: NVIDIA publishes no recipe, so we used our own harness following the protocol in NVIDIA's technical report: static multiple choice without a visual-search tool, full-resolution images, answers matched by letter with a GPT-3.5-Turbo fallback extractor; 191 questions, average of 4 runs.
- GPQA Diamond: official recipe, 198 questions × 8 samples, single run.
- IFBench: official recipe, 300 prompts × 8 samples, strict reward (NVIDIA's metric), single run.
- LiveBench 2.0: overall is the mean of all seven boards. Both models accept image input, so the screenshot-driven Browser board is scored for both.
4. Training Approach
Smaug-Nemotron-3.5 was fine-tuned from the base model to improve agentic coding, tool usage, and long-horizon multi-turn work, using curated multi-turn, tool-using trajectories with the model's reasoning kept in the conversation history. Throughout training we deliberately preserved the base model's original thinking behaviour: training steers the model's actions while keeping its reasoning close to the base, so that the general capabilities of the model are not lost. The reasoning, coding and long-context results above, all at parity with the base, reflect this. Only the language model was adapted; the vision encoder, the vision-language connector and the MTP head are identical to the base. No architectural changes were made.
5. General Capabilities
- Reasoning, coding, instruction-following, long-context and vision benchmarks (GPQA Diamond, LiveCodeBench v6, SciCode, IFBench, AA-LCR, V*) are at parity with the base: the gains are in agentic coding.
6. Deployment
[!Note] Because the architecture is unchanged, Smaug-Nemotron-3.5 runs anywhere the Nemotron 3.5 Super VL BF16 checkpoint runs. Use the base model's deployment recipes with
MODEL=abacusai/smaug-agentic-nemotron.
4× H100 with vLLM (vLLM version: vllm/vllm-openai:v0.32.0), up to 1M context:
export MODEL=abacusai/Smaug-Nemotron-3.5
vllm serve $MODEL \
--served-model-name Smaug-Nemotron-3.5 \
--tensor-parallel-size 4 \
--trust-remote-code \
--kv-cache-dtype fp8 \
--async-scheduling \
--allowed-local-media-path=/ \
--max-num-seqs 32 \
--enable-mamba-shared-prefix-checkpoint \
--limit-mm-per-prompt '{"video": 1, "image": 1}' \
--video-pruning-rate 0.5 \
--media-io-kwargs '{"video": {"fps": 2, "num_frames": 256}}' \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder \
--reasoning-parser nemotron_v3 \
--speculative_config.method mtp \
--speculative_config.num_speculative_tokens 5 \
--default-chat-template-kwargs '{"truncate_history_thinking": false}'
Our evaluations ran on vLLM with --kv-cache-dtype fp8 and without MTP.
7. Model Usage
Sampling behaviour is inherited from the base: temperature 1.0, top_p 0.95, the settings used for every number above (AA-LCR: top_p 1.0, as in NVIDIA's recipe). For image and video input, use the base's vision settings, temperature 0.6, top_p 0.95, as for V* above. For reasoning on/off, image, video, streaming and tool-calling examples, see the base model's API client section; all of them apply unchanged with model = "smaug-agentic-nemotron" (the served model name above).
Coding Agent Framework
The model is trained for multi-turn tool use and works with agent frameworks that speak the OpenAI chat-completions contract.
For multi-turn conversations and tool calls, pass the complete assistant message back to messages, including the reasoning field and tool_calls, not just content. The chat template drops past reasoning by default; serve with truncate_history_thinking: false (as above) to keep it. The agentic coding numbers above were produced with thinking history kept in context.
8. License
Smaug-Nemotron-3.5 is a derivative of NVIDIA Nemotron 3.5 Super VL 120B-A12B and is released under the OpenMDW License Agreement, version 1.1, inherited from the base model. For the base model's bias, explainability, privacy and safety documentation, see the base model card.
This model is not affiliated with or endorsed by NVIDIA.
9. Citation
@misc{abacusai2026smaugnemotron35,
title = {Smaug-Nemotron=3.5},
author = {Abacus.AI},
year = {2026},
note = {Agentic supervised fine-tune of nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16},
url = {https://huggingface.co/abacusai/Smaug-Nemotron-3.5}
}
10. Contact Us
If you have any questions, please reach out at Abacus.AI.