
Nemotron-Labs-Sigma-distill-8B-nfe4 vs Nemotron-Labs-Sigma: Same Weights Lineage, Two Decoding Contracts
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0717 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 72 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 127 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 65 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 146 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 124 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 218 tok/s
The confusing thing about Nemotron-Labs-Sigma-distill-8B-nfe4 is that it is not a better Nemotron-Labs-Sigma-8B. It is a different file with a different contract, produced from a teacher in the same lineage, and the difference between the two is not accuracy — it is how many times the network runs to emit a word. The base 8B was published on 5 October 2026 as a pretrained research model: 8.49B parameters, configurable denoising steps, classifier-free guidance you have to tune by hand, and a card that publishes no scores at all and points to arXiv:2610.02665 instead. The distilled nfe4 arrived on 12 November 2026 with 9.03B parameters, a step count baked into the checkpoint, and the only per-benchmark table NVIDIA has put on a Sigma card. Both are English-only, both cap at 4,096 tokens, both are under the same licence. The choice between them is a choice about how you want the model to spend its compute.
What actually differs, file by file
Both repositories ship the same model_type — nemotron_labs_sigma — and both are loaded the same way, with trust_remote_code and a 40 KB modelling file from the repo. The configs differ in ways that matter at inference and almost not at all in the backbone:
• Backbone — identical shape on both: 34 layers, hidden size 4,096, FFN 14,336, 32 query heads and 8 key-value heads, SwiGLU, RoPE, 131,072-token vocabulary.
• Parameters — 8.49B (8,489,947,136 BF16 plus 2,097,186 F32) for the base, 9.03B (9,026,818,048 BF16 plus the same F32 count) for the distill. The extra 537M is the auxiliary autoregressive readout head that the distill keeps alongside the diffusion head.
• Weights on disk — 16.99 GB against 18.06 GB, four shards each.
• Context — 4,096 tokens for both base and distill. The instruct variant in the same collection raises this to 8,192 and adds a chat template; neither model here has one.
• Steps — the base exposes steps and block_length at call time and its card publishes a per-budget table; the distill accepts no override at all.
• Guidance — the base needs classifier-free guidance (3.0 at 32 and 16 steps, 2.0 at 4) and a tuned score temperature; the distill runs at CFG 1 and score temperature 1, which is to say no guidance and no tuning.
• Sampler — the base is documented for DDIM on a log_snr grid with 64-token blocks by default; the distill uses fresh_x0 with self-conditioning carried between steps and 32-token blocks.
The step count is not a setting on the distilled model
This is the sentence in the model card that reframes the whole comparison: "The number of denoising steps is fixed by training: this variant is trained for NFE=4 only, and the bundled inference code intentionally does not expose an arbitrary step override (other step counts would evaluate untrained phase endpoints)."
Phased DMD distils a trajectory, and a distilled trajectory has phase endpoints at specific noise levels. Ask a four-step student to take eight steps and you are not running it slower for better quality — you are asking it to evaluate at noise levels it never trained on. The card says the output degrades. There is no dial.
The base is the opposite. Its card gives you the table — DDIM-32 with CFG 3 and score temperature 0.3, DDIM-16 with CFG 3 and 0.4, DDIM-8 with CFG 3 and 0.6, DDIM-4 with CFG 2 and 0.8 — and then adds the warning that below roughly 16 steps the pretrained models lose quality noticeably, and that few-step generation should use the distilled variants instead.
Read together, those two statements from the same organisation are a division of labour: the base is the thing you tune, the distill is the thing you do not. If you want a knob, take the base. If you want a constant, take the distill.

The arithmetic that separates them
Put the two contracts in the same units and the reason the distilled file exists becomes obvious.
• Base at its recommended setting — 33 denoiser calls per 64-token block (32 steps plus a readout), which is 1.9 tokens per forward pass.
• Base at 4 steps — 5 calls per 64-token block, 12.8 tokens per pass, but the card warns quality drops noticeably below 16 steps.
• Distill nfe4 — 5 calls per 32-token block, 6.4 tokens per pass, at a quality the card publishes.
• Distill nfe8 — 9 calls per 32-token block, 3.6 tokens per pass, and 4 points better on the six-task mean.
So the distilled nfe4 sits between the base's safe operating point and its unsafe one: three times fewer forward passes per token than the recommended base configuration, with a step count the model was actually trained for. That is the entire premise of the release, and it is a premise about compute, not about intelligence.
What it is not is a measured throughput. The base card reports roughly 120–140 tokens per second for 256 generated tokens on one H100 with DDPM at 32 steps per 64-token block. The distilled card publishes no tokens-per-second figure. The call-count ratios above are arithmetic on published step counts; treat them as such.
Which one has numbers, and what they are worth
Here the two cards diverge sharply, and in a direction that is uncomfortable for the base.
Nemotron-Labs-Sigma-8B publishes no benchmark table. Its evaluation section is two sentences saying results "are reported in the technical report." Everything you can know about its accuracy, you have to take from arXiv:2610.02665 — a paper that covers the 3B and 8B family, both base and distilled, and whose abstract claims the models beat masked diffusion LMs and autoregressive baselines with the largest consistent gains on coding.
The distilled nfe4 card does publish numbers, at ten seeds with two standard errors: GSM8K 86.86 ± 0.28, MATH 61.52 ± 0.20, HumanEval 60.61 ± 1.43, HumanEval+ 57.99 ± 1.20, MBPP 57.50 ± 0.61, MBPP+ 73.70 ± 1.07, six-task mean 66.36 ± 0.46. Nine of the ten seed values are in the range 42 to 51, so the protocol is described precisely enough to reproduce.
The cards as a pair therefore cannot tell you what distillation cost. That is the single most useful number in this comparison and it is missing, because the teacher's card has no table to subtract from. The paper is the only place the claim lives, and its phrasing is directional: roughly 8× acceleration on math and 4× on coding from distribution-matching distillation "with minor accuracy loss." Those multipliers are paper-reported, not card-reported, and they refer to the family's own experiments rather than to a controlled comparison of the exact two checkpoints on the hub today.
One further wrinkle in the distill line itself: the nfe16 sibling is marked pending_ten_seed_six_task_validation. Its three-task mean of 74.50 comes from a single seed-42 run and the card says not to compare it with the other variants' ten-seed means. So "the biggest one is the best" is a reading the card explicitly declines to support.
The third option nobody names
The collection contains a variant that is neither of the two models in this comparison and is arguably the most interesting file in it: Nemotron-Labs-Sigma-distill-spec-8B. It is the same MS-SUSReg-8B teacher, distilled with an additional acceptance objective, and instead of running as a pure few-step sampler it drafts a 32-token block with its diffusion head and verifies it with its own autoregressive head. Each round costs one diffusion call, and accepted tokens are emitted without a further forward pass. The card states that the returned text is exactly the greedy autoregressive output — acceptance changes throughput, never content.
The published acceptance figures, single snapshot: mean accepted length 7.91 on HumanEval, 11.86 on GSM8K, 13.37 on MBPP, bonus token included, bootstrap round excluded. Those are throughput measures, not accuracy measures, and they are not comparable to the nfe4's accuracy table — a distinction the card makes in bold.
Practically, this is the variant to watch if the question is "can a diffusion head accelerate a model without changing what it says." It is not a drop-in replacement for either model in this comparison: it needs generate_speculative, both heads, and self-conditioning switched off.

How to choose in practice
The decision reduces to four checks, and the first one eliminates one side entirely for most readers:
• Are you researching the sampling process itself — guidance schedules, step counts, trajectory steering, classifier-free guidance behaviour? The base is the model, because it is the one with the knobs and the per-budget table.
• Are you building something that generates text at a fixed cost per token and you do not want a tuning surface? The nfe4 distill is the model, because CFG 1 and score temperature 1 is a complete configuration.
• Do you need an assistant? Neither. There is no chat template in either repository, no system prompt, no tool calling, and the base card says plainly that it is not an instruction follower. The instruct variants at 8,192 tokens of context are the ones for that.
• Do you need a benchmark you can cite? Only the distill has one, and it is NVIDIA's own.
None of the four asks about parameter count, which is the tell that this is not a size comparison. The 537M-parameter difference is the price of keeping the autoregressive head in the file, and it buys nothing for the nfe4's pure diffusion path — it is there because the same architecture also underlies the spec variant.
The point of the pair
NVIDIA built one backbone and released it twice: once as something to study and once as something to run. That is a coherent strategy, and it is the reason the two cards read so differently — the base card is written for a researcher who will change the sampler, the distill card for a developer who will not.
What the pair still lacks is the thing that would let you check the trade honestly: a teacher score. Until the base 8B card carries the same six-task table the distill carries, every claim about what the distillation bought is a claim about the paper's family-level experiments rather than about these two files. If NVIDIA adds that table — and the base card's last edit on 16 November suggests the repositories are still being maintained — this comparison becomes a subtraction instead of a reconstruction.
For now, the practical reading is narrow and useful: pick the base if you want to explore how a continuous diffusion LM behaves under different step budgets, pick nfe4 if you want a fixed 6.4 tokens per forward pass at a published accuracy, and skip both if you need a chat model or a context window longer than a long memo. Neither is in the OrcaRouter catalogue today, and neither will be until a serving path exists for a block-diffusion architecture — what OrcaRouter does carry is the 200-plus autoregressive models you would benchmark either of them against, under one key at provider list price with no markup, and automatic failover if you want the experiment isolated from the production path.

