Hy-Memory-30B-A3B vs Nemotron 3.5 Lightning 30B-A3B Base: The Same Three Letters, Two Different Objects

Hy-Memory-30B-A3B and NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Base-BF16 carry identical parameter arithmetic — 30 billion total, 3 billion active — and that is where the resemblance stops being useful. The Hunyuan team's model, open-sourced on October 15, 2026, is a finished product: fully post-trained to execute memory writing inside Mem0, MemOS and Hy-Memory, shipped with an OpenAI-compatible serving recipe and a benchmark table claiming a win over Claude Opus 5. NVIDIA's, released August 11, 2026, is not a product at all. The word "Base" in its name is literal — no supervised fine-tuning, no reinforcement learning, no distillation has been applied. It is pre-trained raw material, published so that you can build your own post-trained model on top of it. What the numbers cannot tell you is whether any of it has gone wrong in the field: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

Two models with the same number in the name, one of which is designed to be used and one of which is designed to be modified. The comparison worth making is not which one scores higher — they have not been evaluated on anything in common, and one of them has no instruction-following behaviour to evaluate. The comparison worth making is about evidence: what each vendor has published, how much of it you could check yourself, and what each one asks of you before it does anything at all.

What each model actually is

Hy-Memory-30B-A3B is a conventional sparse transformer with an unusual job. Its configuration file describes 48 layers, the first a dense feed-forward block and the remaining 47 mixture-of-experts, with 128 experts per block at top-8 routing plus a single shared expert, hidden size 2,048, an expert intermediate width of 768, 32 attention heads against 4 key-value heads with query-key normalisation, and a 262,144-position context. Roughly 60.1 GB across 51 BF16 shards. It was trained in two stages — Memory On-Policy Distillation, then Attributed Memory Reinforcement Learning — and the whole point of the exercise was to replace the general-purpose model that memory frameworks currently call on every turn. Served with vLLM 0.31.0, it returns structured JSON records describing what an agent should remember.

Nemotron 3.5 Lightning 30B-A3B Base is a hybrid machine. NVIDIA interleaves Mamba-2 state-space blocks with mixture-of-experts and attention layers, adds Multi-Token Prediction heads trained in a dedicated continued pre-training phase, and pre-trained the whole thing under an NVFP4 recipe on more than 20 trillion tokens. The point of the hybrid stack is the key-value cache: Mamba-2 blocks keep it from growing the way a conventional transformer's does, which is what lets NVIDIA claim a context length up to 1 million tokens. It ships in BF16, NVFP4 and FP8 builds, with GGUF conversions and quantisations from several third parties. Its licence is OpenMDW-1.1, which covers the weights, the training data and the recipes — a broader grant than a model licence usually is.

The parameter-count coincidence is real but shallow. Both activate about 3B parameters per token; that is the only design decision the two teams made the same way.

That banner is the single most informative element on the page. NVIDIA put it at the top of the base repository on purpose: this checkpoint is not the one most people want, and the vendor says so before you download 30 billion parameters.

The distinction that decides everything: post-trained versus pre-trained

A base checkpoint does not follow instructions. Ask it a question and it continues the text, because that is the only thing pre-training taught it. It has no chat template worth the name, no tool-calling behaviour, no refusal boundary and no idea that a conversation has turns. NVIDIA's card says this outright, and the "Looking for the post-trained model?" banner at the top of the repository points to the sibling checkpoint for anyone who wants something they can talk to.

What a base checkpoint is for is your post-training. Continued pre-training on a domain corpus, or supervised fine-tuning, RL and distillation through NeMo RL, NeMo Gym and Megatron-LM — the toolchain NVIDIA names and ships recipes for. If you have a data advantage and an opinion about how a model should behave, a base checkpoint is the starting point that lets you act on both. If you do not, it is a 30-billion-parameter paperweight.

Hy-Memory-30B-A3B is at the opposite end of that pipeline. It has already been through post-training, twice over, and the entire value proposition is that you do not have to do that work — you take the schema the card gives you, point it at your harness, and it writes memory. It is narrower than Nemotron by an enormous margin and finished in a way Nemotron deliberately is not.

Specs and packaging, one line per dimension

• Total / active parameters — 30B / ~3B for both. This is the only row where they agree.

• Architecture — Hy-Memory-30B-A3B is a sparse transformer with one dense block and 47 MoE blocks, 128 experts at top-8 plus a shared expert; Nemotron 3.5 Lightning 30B-A3B Base is a Mamba-2 hybrid, interleaving state-space, MoE and attention layers with Multi-Token Prediction heads.

• Context — 262,144 positions for Hy-Memory-30B-A3B from its config; up to 1M tokens for Nemotron, with 256K as the documented practical figure on a single H100.

• Post-training — two stages for Hy-Memory-30B-A3B (on-policy distillation, then attributed RL); none at all for the Nemotron base checkpoint.

• Weight footprint — 60.1 GB BF16 for Hy-Memory-30B-A3B with no other precision published vs BF16, NVFP4 and FP8 builds from NVIDIA plus GGUF from the community.

• Licence — Apache 2.0 vs OpenMDW-1.1, which additionally covers training data and recipes.

• Serving path — vLLM 0.31.0 behind an OpenAI-compatible endpoint for Hy-Memory-30B-A3B; post-train first, then serve, for Nemotron.

• Hosted endpoint — neither model is deployed by a hosted inference provider. Both are self-hosting propositions.

Evidence: NVIDIA publishes a recipe, Hunyuan publishes a table

This is the sharpest difference between the two releases, and it is not about the numbers themselves.

NVIDIA publishes its evaluation methodology in a way you can act on. The Lightning base card states that its accuracy figures were measured under a consistent harness — NeMo Gym and the NeMo Evaluator SDK, wrapping an open-source container of LM Evaluation Harness, with RULER run through NeMo Skills — and then says the thing vendors rarely say: these numbers may differ from other vendors' self-reported figures. It links reproducibility recipes. The numbers are still NVIDIA's own run, but the harness is named, containerised and downloadable, which means the distance between "NVIDIA measured this" and "I measured this" is a few hours of work rather than a research project.

The Hy-Memory release is harder to check in two distinct ways. The evaluation design is sound on its own terms — it changes only the memory writer and holds the harness, prompts, retrieval configuration, answer model and judge fixed, which is the correct way to isolate a writer's contribution. But every benchmark in the main matrix runs on a fixed 20% subsample rather than the full set, and the headline margin is thin: 55.49 macro average against 54.62 for Claude Opus 5, a 0.87-point gap, with the MemOS harness decided by 0.01 of a point. The second obstacle is hardware. Reproducing NVIDIA's base numbers means loading a checkpoint whose quantised builds exist; reproducing Hy-Memory's means loading 60 GB of BF16 with no smaller alternative published.

Neither set of figures has been independently reproduced. The difference is how plausible it is that you could be the one to do it.

Put the two primary sources side by side and the contrast is stark: The Hy-Memory evaluation lives in a paper with an architecture description and a training method, and NVIDIA's lives in a model card with a named, downloadable harness. Both are the vendor marking its own homework. Only one of them hands you the ruler.

Two different reasons to bring a router into it

Both models are self-hosted, and neither Hy-Memory-30B-A3B nor Nemotron 3.5 Lightning 30B-A3B Base is in OrcaRouter's catalogue — we host neither of them. The route to both is the weights and your own GPUs. But the router question lands differently on each side of this comparison, which is worth walking through because it is the practical shape of the choice.

If you take the Nemotron path, you are post-training a checkpoint. That produces a model that exists only on your infrastructure, and the classic problem with a custom checkpoint is that it has nowhere to sit: it is not in any catalogue, so every application that wants it needs bespoke plumbing, and every fallback path when it goes down needs to be built by hand. A router solves precisely that — one API key across 200+ models, with your own endpoint registered alongside them and automatic failover when it degrades. Model fusion is the other piece worth knowing about here: it runs a panel of models together on one call, which is how you would compare a fresh post-train against a strong baseline without writing a harness for each.

If you take the Hy-Memory path, the writer is one call in a much larger loop. A memory harness invokes the writer, then invokes an answer model to read the bank back and respond — and that second call is where a hosted model normally sits. OrcaRouter passes provider list price through with 0% markup, so a vendor price cut on the answer model is live on our side the same day, and the routing DSL composes the self-hosted writer and the hosted reader into a single call. You do not need both halves to come from the same place, and you do not need a second contract to add the one you are missing.

Which one is the right starting point

If what you want is a model that does something on the day you download it — specifically, that makes an agent's memory better without you training anything — Hy-Memory-30B-A3B is the only one of these two that answers. It is also the one you cannot test cheaply, and its headline claim is the one with the thinnest margin. Go in expecting to spend on hardware before you learn anything.

If what you want is a foundation you can shape, Nemotron 3.5 Lightning 30B-A3B Base is a serious piece of published work: a hybrid architecture designed for long context and fast inference, released with its training recipes and an evaluation harness you can re-run. It will not answer a question until you have post-trained it, and that is the deal rather than a defect.

The name collision will keep putting these two in the same sentence. The parameter counts will keep matching. Nothing else about them will, and the sooner a reader separates "30B-A3B" from "the same kind of model," the sooner the comparison becomes useful.

The download figure is worth leaving on screen. A post-trained specialist released yesterday with no quantised build, no hosted endpoint and no announcement has zero downloads and nine likes — and NVIDIA's pre-trained base checkpoint, published two months earlier, sits in the tens of thousands. That gap is not a quality judgement. It is a measurement of how hard each team made it to try the thing.