Reply To: Debt Rattle November 16 2026
Home › Forums › The Automatic Earth Forum › Debt Rattle November 16 2026 › Reply To: Debt Rattle November 16 2026
The following has a lot of links and will be loaded in multiple parts.
I want you to tell me about … Nemotron 3 Ultra and that it would be competing directly with Claude Opus…
Since it would be about technical info leave me with links to open and read
Nemotron 3 Ultra is already out. NVIDIA released it on June 4, 2026, after pre-announcing it at Computex, on Hugging Face, ModelScope, OpenRouter and build.nvidia.com. It competes with Opus on cost, speed and openness, not on raw capability.
**What it is**
– An open model built for long-running agents, with weights, data and recipes all published.
– 550B total and 55B active parameters, a hybrid Mamba-Transformer MoE with latent MoE routing, multi-token prediction and a 1M-token context window.
– It is the largest and most capable model in the Nemotron 3 family.
– It is released under OpenMDW-1.1, with an NVFP4 checkpoint for Blackwell, Hopper and Ampere.
– NVIDIA claims about 5x the throughput of other open models in its class.
**How it compares with Claude Opus**
– It is a leading US open-weights model. On Artificial Analysis’s Intelligence Index it scored 47.7 using the NVFP4 weights. The comparison sites below don’t all use the same index or benchmark mix, so the absolute scores differ between them.
– It trails Opus on coding and agentic work. One comparison against Opus 4.8 has Opus ahead on 20 of 21 shared benchmarks, such as SWE-bench Verified at 88.6% vs 69.7%. Another has Opus 5 at 96% on SWE-bench Verified vs 71.9% for Nemotron. I didn’t find a comparison against Opus 5.5 specifically, and none of this is NVIDIA’s own claim.
– It wins on price and speed. It lists at $0.50 input and $2.20 output per 1M tokens, versus $5 and $25 for Opus 4.8. Artificial Analysis also measured it at 223 tokens/s against 55 for Opus 4.5.
– It has no image input. Opus 4.5 does, and Nemotron doesn’t.
– Because the weights are open, you can self-host and fine-tune it, which you can’t do with Opus.
NVIDIA’s own framing is “leading accuracy among open models” rather than a match for Opus. It makes sense where cost, data control or customization matter more than the top coding and agent scores.
————–







Sorry, the comment form is closed at this time.