Nemotron 3 Ultra
A large language model released by NVIDIA belonging to the Nemotron series. It is available at no cost through multiple inference services and supports code understanding and generation capabilities.
Timeline (last 90 days)
Indonesian-language field accuracy measurements: Nemotron 3 Ultra 84.8%, Nemotron 3 Super 70.7%, Nemotron 3 Lightning 59.6%, with reference labels drafted by an agent and not human-reviewed.
Awas uses Nemotron 3 Ultra on Nebius Token Factory to extract structured fields from ads and uses Nemotron 3 Super to propose evidence rows.
Among the 38 free models, only 4 are suitable for writing code: GLM5.3, DeepSeek 4.1 Flash, Nemotron 3 Ultra, and Poolside Laguna XS 2.1.
In the real leaderboard data, HY3 (free) scores 82.9 and Nemotron 3 Ultra (free) scores 81.7, both provided by opencode-zen.
NVIDIA used knowledge distillation to compress the capabilities of the Nemotron 3 family (including Ultra) into Nemotron 3.5 Super VL.
Some users pointed out that Nvidia's 550B-parameter Nemotron 3 Ultra (released in June) is only on par with Qwen3.5-27B, which was released in February.
Among the 80 free models, 16 are current-generation chat or reasoning models, 32 are NVIDIA's own models (including Nemotron 3 Ultra and Super, safety classifiers, embedding models, and a 340B reward model), and 26 are general-purpose chat models from 2023 to 2025 (such as Llama 2 70B, Mixtral 8x22B, Yi Large, and DBRX).
Six of the world's top eleven open-weight models are on NVIDIA's free endpoint, including the second-ranked GLM-5.3 and the fifth-ranked DeepSeek V4.1 Flash; NVIDIA's own Nemotron 3 Ultra ranks eleventh on that index.
Nemotron 3 Ultra scored 71.46666666666667 on this benchmark, is called for free via opencode, and is labeled by the provider as deepseek.
Nemotron 3 Ultra (via opencode) scored 43.49 on this benchmark.
9 more timeline events
All Facts (19)
Indonesian-language field accuracy measurements: Nemotron 3 Ultra 84.8%, Nemotron 3 Super 70.7%, Nemotron 3 Lightning 59.6%, with reference labels drafted by an agent and not human-reviewed.
50%UnverifiedAwas uses Nemotron 3 Ultra on Nebius Token Factory to extract structured fields from ads and uses Nemotron 3 Super to propose evidence rows.
50%UnverifiedAmong the 38 free models, only 4 are suitable for writing code: GLM5.3, DeepSeek 4.1 Flash, Nemotron 3 Ultra, and Poolside Laguna XS 2.1.
50%UnverifiedIn the real leaderboard data, HY3 (free) scores 82.9 and Nemotron 3 Ultra (free) scores 81.7, both provided by opencode-zen.
50%UnverifiedNVIDIA used knowledge distillation to compress the capabilities of the Nemotron 3 family (including Ultra) into Nemotron 3.5 Super VL.
50%UnverifiedSome users pointed out that Nvidia's 550B-parameter Nemotron 3 Ultra (released in June) is only on par with Qwen3.5-27B, which was released in February.
50%UnverifiedAmong the 80 free models, 16 are current-generation chat or reasoning models, 32 are NVIDIA's own models (including Nemotron 3 Ultra and Super, safety classifiers, embedding models, and a 340B reward model), and 26 are general-purpose chat models from 2023 to 2025 (such as Llama 2 70B, Mixtral 8x22B, Yi Large, and DBRX).
50%UnverifiedSix of the world's top eleven open-weight models are on NVIDIA's free endpoint, including the second-ranked GLM-5.3 and the fifth-ranked DeepSeek V4.1 Flash; NVIDIA's own Nemotron 3 Ultra ranks eleventh on that index.
50%UnverifiedNemotron 3 Ultra scored 71.46666666666667 on this benchmark, is called for free via opencode, and is labeled by the provider as deepseek.
50%UnverifiedNemotron 3 Ultra (via opencode) scored 43.49 on this benchmark.
50%UnverifiedSolar Masonry E4 scored 24 on the Artificial Analysis Intelligence Index, the highest score in the 3-billion active parameter tier, trailing NVIDIA's Nemotron 3 Ultra by only one point.
50%UnverifiedThe system features a separate high-compute stage responsible for selecting the final submitted answer for each problem.
50%UnverifiedThe system trained based on Nemotron 3 Ultra achieved a score of 30 out of 42 at IMO 2026, reaching the gold medal threshold.
50%UnverifiedThe system simultaneously invokes three Nemotron 3 Ultra checkpoints—a general-purpose version plus two post-training specialists—to drive an iterative search process.
50%UnverifiedStarting from Nemotron 3 Ultra, the team trained two expert checkpoints specifically targeting math olympiad problems through supervised fine-tuning (SFT) and reinforcement learning (RL).
50%UnverifiedThe system operates entirely at the natural language level, without relying on formal proof verifiers, calling external tools, or having internet access.
50%UnverifiedThe test-time compute pipeline employs an iterative loop mechanism of generation, verification, and refinement.
50%UnverifiedAfter full-stack optimization, the unit serving cost has been reduced to approximately 40% of the pre-optimization level.
50%UnverifiedThrough full-stack NIM optimization, the concurrent serving capacity of the Nemotron 3 Ultra model has improved by 2.5x.
50%