A generated title card reading Nemotron-Labs-Sigma-distill-8B-nfe4 vs Nemotron 3.5 Lightning 30B A3B, subtitled that one is a five-download diffusion research checkpoint and the other a 962,098-download production MoE, with a footer stating that both sets of figures are vendor-reported and no benchmark is shared between them.
Guides & Insights

Nemotron-Labs-Sigma-distill-8B-nfe4 vs Nemotron 3.5 Lightning 30B A3B: Two NVIDIA Releases That Never Compete

Author

Gideon Frost

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Put Nemotron-Labs-Sigma-distill-8B-nfe4 and Nemotron 3.5 Lightning 30B A3B in the same sentence and the instinct is to ask which is smarter. That question has no answer, and not because the data is thin — because the two models share no benchmark, no context length, no serving stack and no release process. Nemotron-Labs-Sigma-distill-8B-nfe4 is an 8B continuous diffusion model that NVIDIA published on 12 November 2026 with no announcement and has been downloaded five times. Nemotron 3.5 Lightning 30B A3B is a 30B-total, 3B-active hybrid Mamba-2 and attention mixture-of-experts that shipped in August with a model card, a build.nvidia.com endpoint and 962,098 downloads last month. The useful comparison is not a scoreline. It is which one you can actually put in front of a user, and for what.

The two cards describe different products, not different tiers

Read the use-case sections side by side and the gap stops being about capability. Sigma-distill's card says it is for "latency-sensitive generation research, speculative and parallel decoding studies, and deployment-oriented experimentation," and states that the model "runs offline on the user's own hardware; NVIDIA hosts no service." Lightning's card points at an OpenAI-compatible endpoint at integrate.api.nvidia.com, a vLLM container tag, a recommended sampling pair of temperature 1.0 and top_p 0.95, a thinking-mode toggle in the chat template, and three separate speculative-decoding draft checkpoints.

One is a file. The other is a service with a file behind it. That difference drives almost everything else in this piece:

• Parameters — 8B dense (9.03B in the safetensors census once the auxiliary readout head is counted) vs 30B total with 3B active per token.

• Architecture — Ministral-style decoder with a 16-dimensional continuous diffusion head vs interleaved Mamba-2, MoE and selective-attention layers with multi-token-prediction heads.

• Context — 4,096 tokens for prompt and output combined vs up to 1,000,000, with the caveat that a single H100 only reaches about 256,000 and the full million needs eight H100s, a GB200 or a B200.

• Languages — English only vs English plus coding languages, Spanish, French, German, Italian and Japanese, with Chinese in post-training.

• Release shape — card and weights only, no vendor endpoint, no announcement vs a hosted endpoint, a BF16 reference build, an NVFP4 build, and three draft models for speculative decoding.

• Downloads in the last thirty days — five vs 962,098 on the BF16 repository alone, before counting 901,381 on the NVFP4 build.

• Licence — NVIDIA Nemotron Open Model License for both, though Lightning's card names it OpenMDW-1.1.

There is no shared benchmark, so there is no scoreline

This is the part most comparison pages would paper over. Sigma-distill's card reports GSM8K, MATH (Minerva), HumanEval, HumanEval+, MBPP and MBPP+ at ten seeds with two standard errors, plus a six-task mean of 66.36. Lightning's card reports MMLU Pro, GPQA Diamond, HLE, SciCode, SWE-bench Verified, SWE-bench Multilingual, Terminal-Bench 2.1, PinchBench, BrowseComp, τ³-bench Banking, GDPval-AA-V2, IFBench, AA-LCR and AA-Omniscience. The intersection is the empty set. Not one benchmark appears on both cards.

That is not an accident of reporting. The two evaluations test different things on purpose. Sigma's six tasks are short-answer reasoning and code completion at a few hundred tokens; Lightning's are agentic software engineering, long-context retrieval and professional knowledge. A model with a 4,096-token window and no chat template cannot sit the SWE-bench harness, and a model trained to finish a function body is not being asked GPQA questions.

So the only honest line to draw is this: on its own six tasks, Sigma-distill nfe4 averages 66.36 at ten seeds, which is a real number under a real protocol, and it says nothing about how it would do on anything Lightning was measured on. Anyone who tells you one beats the other is comparing nothing.

A generated chart showing that the two model cards share no benchmark: on the left, the six Nemotron-Labs-Sigma-distill-8B-nfe4 tasks with ten-seed accuracies of GSM8K 86.86, MATH 61.52, HumanEval 60.61, HumanEval+ 57.99, MBPP 57.50 and MBPP+ 73.70, mean 66.36; on the right, the Nemotron 3.5 Lightning 30B A3B evaluations including MMLU Pro 81.94, GPQA Diamond 75.44, SWE-bench Verified 51.56 and Terminal-Bench 2.1 24.58, with a divider marking that no task appears on both sides.

Speed is where the comparison gets interesting — and where it stays unresolved

Both models are fast in ways that are hard to put on one axis, and this is the one place the matchup has real content.

Sigma-distill generates 32 tokens per block in exactly 4 denoising evaluations plus one readout. Five forward passes, 32 tokens: 6.4 tokens per forward pass. Because the step count is fixed by training, that ratio does not change with prompt length or output length — it is a constant, and the card's inference code refuses to let you override it. The parent model's card reports 120–140 tokens per second on one H100 at 32 steps per 64-token block; the distilled variant publishes no tokens-per-second figure at all, so the 6.4×-fewer-passes arithmetic is not a measured throughput.

Lightning takes the other route. It is a 3B-active MoE, so each forward pass is cheap, and NVIDIA ships DSpark as the recommended speculative-decoding default with DFlash and its own multi-token-prediction heads as alternatives. Its throughput numbers are also not on the card in tokens per second — but its downloads and its serving recipe tell you it has been run in production, which Sigma has not.

The distinction that matters: Sigma's speedup is structural and constant per token, while Lightning's is speculative and workload-dependent. A fixed 5-calls-per-32-tokens contract is predictable in a way that draft acceptance rates never are. It is also unmeasurable today, because nobody has published a single throughput number for the nfe4 checkpoint.

What each one costs you to run

Sigma-distill is 18.06 GB of bfloat16 weights across four shards, loaded through PyTorch eager code from the repository itself with trust_remote_code, on a CUDA GPU, with no paged attention, no continuous batching and no quantised build in the collection. There is no per-token price because there is no endpoint: NVIDIA's card states the model runs on your hardware and that NVIDIA hosts no service.

Lightning is the opposite shape. BF16 weights are 31.6B parameters; NVIDIA's own production advice is to use the NVFP4 build instead, and its card ships a vLLM recipe (container tag vllm/vllm-openai:v0.27.1) with FlashInfer Mamba backend, FP16 SSM cache and prefix caching, plus an SGLang image for Hopper and Blackwell. There is no list price on the card or on build.nvidia.com — the endpoint is under NVIDIA's API trial terms — but third-party platforms list the model at roughly $0.05 per million input tokens and $0.20 per million output tokens. That figure is a marketplace observation, not a vendor price.

Neither model is in the OrcaRouter catalogue, and it would be dishonest to imply otherwise. What is available at OrcaRouter is the surrounding comparison set: 200-plus models behind one key, provider list price passed through with no markup, so a price cut upstream shows up on the same day rather than in a quarterly rate-card revision. If the reason you are reading this is that you need to pick a baseline to test Sigma against, that is where the baseline comes from.

Pick by envelope, not by score

Four questions sort this in under a minute, and none of them involve benchmark accuracy:

• Does the task fit in 4,096 tokens? If not, Sigma-distill is out and the question is over. Lightning's million-token window — 256K on a single H100 — is three orders of magnitude larger.

• Does the task need conversation, tools or structured output? Sigma-distill has none of it: no chat template, no system prompt, no JSON mode, stop strings and a token budget as the only stopping mechanisms. Lightning ships a chat template with a thinking toggle.

• Do you need to serve more than one request at a time? Sigma has no serving stack — the supported path is a Python script on one GPU. Lightning has two engine integrations and a published parallelism recipe.

• Is your interest in the decoding mechanism itself — block diffusion, few-step distillation, self-speculation? Then Sigma-distill is the only one of the two that is interesting, and Lightning is the wrong comparison entirely.

Everything else follows from those four. A team evaluating Sigma for a product is really asking whether a research checkpoint can be turned into a service; a team evaluating Lightning is asking whether a hybrid MoE at 3B active fits a budget. Those are different conversations.

Where the matchup would become real

Two things would turn this page into a genuine head-to-head, and neither exists yet.

The first is a shared benchmark. If Sigma ever gets an MMLU Pro or GPQA number, or Lightning ever gets a GSM8K or HumanEval number at a comparable prompt format, there is something to compare. Until then the honest position is the one above: two disjoint evaluation suites reported by the same vendor, on models that cannot sit each other's tests.

The second is a serving path. Sigma-distill is currently a repository you clone and a script you run; the SGLang pull requests that mention Nemotron diffusion target the older Nemotron-Labs-Diffusion-8B lineage and are still open as of 17 November, and vLLM's registry has no nemotron_labs_sigma architecture at all. The moment a block-diffusion model with a fixed step count gets paged attention and batching, its 6.4-tokens-per-pass advantage becomes a number you can put next to a speculative-decoding acceptance rate — and that comparison would be worth publishing.

Screenshot of the Hugging Face model card for NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16, showing the model description as a hybrid Mamba-2, mixture-of-experts and attention model with 30B total and 3B active parameters, the release date of 08/11/2026, the OpenMDW-1.1 licence, and the evaluation table beginning with MMLU Pro at 81.94 and GPQA Diamond at 75.44.

Until then the practical reading is that NVIDIA released two models for two audiences in the same autumn and only one of them was a product. Lightning is the one you can call today, with the cache behaviour, the context window and the engine support that a service implies. Sigma-distill is the one you download if you care about whether continuous diffusion at 8B can hold 66.36 on six tasks with five forward passes per 32 tokens — a question about decoding, not about capability, and one that no amount of comparison against a 30B MoE will answer.

A generated two-column scoreboard titled Nemotron-Labs-Sigma-distill-8B-nfe4 vs Nemotron 3.5 Lightning 30B A3B, contrasting context of 4,096 tokens against up to 1,000,000, English only against seven languages, five downloads against 962,098, no vendor endpoint against a hosted OpenAI-compatible endpoint, no engine support against vLLM and SGLang recipes, and a fixed 6.4 tokens per forward pass against speculative decoding, with a footer noting that no benchmark is shared between the two cards.