NVIDIA has released Nemotron 3 Super, a 120B (12B active) open weights reasoning model that scores 36 on the Artificial Analysis Intelligence Index with a hybrid Mamba-Transformer MoE architecture We were given access to this model ahead of launch and evaluated it across intelligence, openness, and inference efficiency. Key takeaways ➤ Combines high openness with strong intelligence: Nemotron 3 Super performs strongly for its size and is substantially more intelligent than any other model with comparable openness ➤ Nemotron 3 Super scored 36 on the Artificial Analysis Intelligence Index, +17 points ahead of the previous Super release and +12 points from Nemotron 3 Nano. Compared to models in a similar size category, this places it ahead of gpt-oss-120b (33), but behind the recently-released Qwen3.5 122B A10B (42). ➤ Focused on efficient intelligence: we found Nemotron 3 Super to have higher intelligence than gpt-oss-120b while enabling ~10% higher throughput per GPU in a simple but realistic load test ➤ Supported today for fast serverless inference: providers including @DeepInfra and @LightningAI are serving this model at launch with speeds of up to 484 tokens per second Model details 📝 Nemotron 3 Super has 120.6B total and 12.7B active parameters, along with a 1 million token context window and hybrid reasoning support. It is published with open weights and a permissive license, alongside open training data and methodology disclosure 📐 The model has several design features enabling efficient inference, including using hybrid Mamba-Transformer and LatentMoE architectures, multi-token prediction, and NVFP4 quantized weights 🎯 NVIDIA pre-trained Nemotron 3 Super in (mostly) NVFP4 precision, but moved to BF16 for post-training. Our evaluation scores use the BF16 weights 🧠 We benchmarked Nemotron 3 Super in its highest-effort reasoning mode ("regular"), the most capable of the model's three inference modes (reasoning-off, low-effort, and regular)

Alibaba has released 4 new Qwen3.5 models from 0.8B to 9B. The 9B (Reasoning, 32 on the Intelligence Index) is the most intelligent model under 10B parameters, and the 4B (Reasoning, 27) the most intelligent under 5B, but both use 200M+ output tokens to run the Intelligence Index @Alibaba_Qwen has expanded the Qwen3.5 family with four smaller dense models: the 9B (Reasoning, 32 on the Intelligence Index), 4B (Reasoning, 27), 2B (Reasoning, 16), and 0.8B (Reasoning, 9). These complement the larger 397B, 27B, 122B A10B, and 35B A3B models released earlier this month. All models are Apache 2.0 licensed, support 262K context, include native vision support, and use the same unified thinking/non-thinking hybrid approach as the rest of the Qwen3.5 family Key benchmarking results for the reasoning variants: ➤ The 9B and 4B are the most intelligent models at their respective size classes, ahead of all other models under 10B parameters. Qwen3.5 9B (32) scores roughly double the next closest models under 10B: Falcon-H1R-7B (16) and NVIDIA Nemotron Nano 9B V2 (Reasoning, 15). Qwen3.5 4B (27) outscores all of these despite having roughly half the parameters. All four of the small Qwen3.5 models are on the Pareto frontier of the Intelligence vs. Total Parameters chart ➤ The Qwen3.5 generation represents a material intelligence uplift over Qwen3 across all sub-10B model sizes, with larger gains at higher total parameter counts. Comparing reasoning variants: Qwen3.5 9B (32) is 15 points ahead of Qwen3 VL 8B (17), the 4B (27) gains 9 points over Qwen3 4B 2507 (18), the 2B (16) is 3 points ahead of Qwen3 1.7B (estimated 13), and the 0.8B (9) gains 2.5 points over Qwen3 0.6B (6.5). ➤ All four models use 230-390M output tokens to run the Intelligence Index, significantly more than both larger Qwen3.5 siblings and Qwen3 predecessors. Qwen3.5 2B used ~390M output tokens, 4B used ~240M, 0.8B used ~230M, and 9B used ~260M. For context, the much larger Qwen3.5 27B used 98M and the 397B flagship used 86M. These token counts also exceed most frontier models: Gemini 3.1 Pro Preview (57M), GPT-5.2 (xhigh, 130M), and GLM-5 Reasoning (109M) ➤ AA-Omniscience is a relative weakness, with hallucination rates of 80-82% for the 4B and 9B. Qwen3.5 4B scores -57 on AA-Omniscience with a hallucination rate of 80% and accuracy of 12.8%. Qwen3.5 9B scores -56 with 82% hallucination and 14.7% accuracy. These are marginally better than their Qwen3 predecessors (Qwen3 4B 2507: -61, 84% hallucination, 12.7% accuracy), with the improvement driven primarily by lower hallucination rates rather than higher accuracy. ➤ The Qwen3.5 sub-10B models combine high intelligence with native vision at a scale previously unavailable. On MMMU-Pro (multimodal reasoning), Qwen3.5 9B scores 69.2% and 4B scores 65.4%, ahead of Qwen3 VL 8B (56.6%), Qwen3 VL 4B (52.0%), and Ministral 3 8B (46.0%). The Qwen3.5 0.8B scores 25.8%, which is notable for a sub-1B model Other information: ➤ Context window: 262K tokens ➤ License: Apache 2.0 ➤ Quantization: Native weights are BF16. Alibaba has not released first-party GPTQ-Int4 quantizations for these small models, though they have for the larger models in the Qwen3.5 family released earlier (27B, 35B-A3B, 122B-A10B, 397B-A17B). In 4-bit quantization all four models are accessible on consumer hardware ➤ Availability: At time of publishing, there are no first-party or third-party serverless APIs hosting these models