This links to the nvfp4 version of the model, so they only compare it to the bf16 in this model card. If you're looking for other similar model comparisons they are in the model card of the bf16 version here [1].
While it looks "behind" the qwen equivalent model on most benchmarks, a few personal notes:
- nemotron models feel to me a bit less benchmaxxed / "stubborn". That means that they generalise a bit better, or can be tasked to solve similar but not quite identical task types to the training data (something that's hard to do w/ qwen/ds models)
- nemotron series are also open training (w/ open training recipes and some training data public)
- nvda will have an incentive to continue this kind of releases, even if other parties slowly abandon the open release of models. Whatever other incentives 3rd party labs have (i.e. meta, goog w/ gemma, the chinese labs that IPOd, etc) nvda will always want to sell hardware so their incentive to keep pushing open models is evident and will likely continue "forever".
this "DFlash" model should be used together with one of the previous two "for lower-latency speculative decoding deployments tuned for low-concurrency data center and workstation workflows".
Crazy to see how well Qwen3.6 35b-a3b is holding up, sure it is ~20% larger but it's scores are also ~20% higher with the same number of active params (excluding the IFBench).
Hopefully Qwen follows up their 3.8 launch with a new 35b-a3b
4 months of progress. Qwen 3.8 Max is, depending on who you ask, what you're doing, standard disclaimers, really really close to Opus 4.8. That is a huge dump from 3.6 Max. Presumably improvements in their training librar(y/ies), training improvements, etc will "trickle down" to Qwen 3.8. We are using 3.6 35B internally and it is in most cases better than Sonnet 4.6 which already makes it "good enough" for most data enrichment tasks. I would be a lot more excited about a MoE Qwen 3.8 35B-A3B than the "announced" 27B dense model. I have 50 million records I need to keep up to date, an A3B MoE is waaaay more performant than a 27B dense model.
Interesting to see in the Agentic Coding Benchmarks that the codex harness is so demonstrably lower than any of the other harnesses, even Claude Code another lab harness. Does anyone know why that would be?
You might look at this and and be a bit disappointed by the performance against qwen and gemma models - but this is an entirely open source training pipeline, this is quite impressive and I don't think another model this performant exists with fully open source data and recipes alongside the weights.
Nvidia just throwing something "for peasants" to stay relevant. Where is competition spirit? More importantly why Nvidia is gatekeeping computing for everyday people?
I find these releases are bad taste.
Make 1TB DGX priced affordably, not some crap model for people to waste time on.
imo a 512GB DGX cluster for $20k is not a bad price for an enterprise that wants to do on prem LLMs. Gatekeeping is a weird way to put “running a business”
While it looks "behind" the qwen equivalent model on most benchmarks, a few personal notes:
- nemotron models feel to me a bit less benchmaxxed / "stubborn". That means that they generalise a bit better, or can be tasked to solve similar but not quite identical task types to the training data (something that's hard to do w/ qwen/ds models)
- nemotron series are also open training (w/ open training recipes and some training data public)
- nvda will have an incentive to continue this kind of releases, even if other parties slowly abandon the open release of models. Whatever other incentives 3rd party labs have (i.e. meta, goog w/ gemma, the chinese labs that IPOd, etc) nvda will always want to sell hardware so their incentive to keep pushing open models is evident and will likely continue "forever".
[1] - https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-...
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-...
to be used for further training/fine-tuning.
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-...
main model.
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-...
quantized version of the previous.
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-...
this "DFlash" model should be used together with one of the previous two "for lower-latency speculative decoding deployments tuned for low-concurrency data center and workstation workflows".
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-...
like DFlash, the previous model above, but optimized for DGX Spark.
I actually copied the link from NVIDIA's Technical Blog post:
- https://developer.nvidia.com/blog/nvidia-nemotron-3-5-lightn...
You can also try the model via a free API endpoint from Openrouter, would be interesting to see if it's the BF16 or NVFP4 version:
- https://openrouter.ai/nvidia/nemotron-3.5-lightning:free
This! It's literally in their best interest for open-weights models to succeed
I used LM Studio and this model file: https://lmstudio.ai/models/nvidia/nemotron-3.5-lightning
I got a pretty solid (and reasonably fast) result for this simply coding agent activity (with simonw/llm-coding-agent):
https://gist.github.com/simonw/a8741f79280cedc86bcb8d90edfb4... - took just under a minute.Hopefully Qwen follows up their 3.8 launch with a new 35b-a3b
If you're already selling shovels, you might as well provide maps to the gold mine with every purchase.
https://github.com/NVIDIA-NeMo/Nemotron/blob/main/docs/nemot...
Changing the SSM backbone is a major revision, not a dot release. At least for us.
I find these releases are bad taste.
Make 1TB DGX priced affordably, not some crap model for people to waste time on.