One-line takeaway
Releasing 14 variants of one weight family lets you pick by scenario, but it also means every score must be matched to its configuration: the 71.16 for base and the 70.34 for a distilled variant are not the same thing.
Indicator summary
| Indicator | Value | Convention and source |
|---|---|---|
| Checkpoints | 14 | Hugging Face collection |
| Parameter tiers | 3B and 8B | Model cards |
| Variant classes | base, instruct, few-step distilled, self-speculative | Collection grouping |
| Generation | Continuous diffusion, denoising in a 16-dim embedding space | Technical report |
| Block size | 64 tokens per block | Technical report |
| Training start | Warm start from a ~1T-token autoregressive checkpoint | Technical report |
| Diffusion training | About 0.3T tokens | Technical report |
| License | NVIDIA Nemotron Open Model License | Model card |
| CPU support | None | Model card |
| Quantized variants | None | Model card |
How the 14 break down
| Class | Count | Notes |
|---|---|---|
| base | 2 | One 3B, one 8B, no instruction tuning |
| instruct | 2 | One 3B, one 8B; the 8B instruct has 8192 context |
| few-step distilled | 8 | NFE (denoising steps) 2 / 4 / 8 / 16 across both tiers |
| self-speculative | 2 | One per size, for accelerated decoding |
NFE is the number of denoising steps in a diffusion model. Fewer steps mean faster generation per sample and possibly lower quality. Offering four settings hands the speed-quality curve to the user.
Six-task averages
| Model | Six-task average | Note |
|---|---|---|
| Sigma-8B base | 71.16 | The headline model |
| NLD-8B | 70.22 | Sibling comparison |
| Qwen3-8B | 68.91 | Autoregressive comparison |
| Dream-7B | 58.74 | Discrete diffusion comparison |
| LLaDA-8B | 41.78 | Discrete diffusion comparison |
The value of this table is that all five ran the same six tasks, so the gap between 71.16 and 68.91 is meaningful. Note these are base scores without instruction tuning, not interchangeable with instruct results.
8B instruct on key benchmarks
| Benchmark | Sigma-8B instruct |
|---|---|
| GSM8K | 91.96 |
| AIME25 | 17.00 |
This pair best describes the model's character: near-saturated on elementary math, clearly struggling on competition math. The explanation is not hard — diffusion-style generation is good at refining patterned computation within a block, while problems that need long chains of leaps run into both the 4K context and block-level generation.
A second reference point is the distilled checkpoint card: distill-8B-nfe8 scores 70.34 on the six tasks, 88.44 on GSM8K, 65.18 on HumanEval and 62.60 on MBPP, noted as results over 10 random seeds. Against the base 71.16, stepping down the denoising count costs less than a point on the average. Whether hard problems such as AIME stay as stable is not published.
Speed
| Indicator | Value | Condition |
|---|---|---|
| Single-GPU throughput | About 120–140 tokens/s | H100, bf16 |
| Average self-speculative speedup | 3.46x | 8B, GB200 with SGLang |
| Peak self-speculative speedup | 4.20x | Same |
At 120–140 tokens/s for an 8B model, this is mid-to-upper range, but measured on an H100 in bf16 without speculative decoding. The speedups are relative to an autoregressive baseline on different hardware (GB200), so the two figures cannot be multiplied to predict a combined number.
Why 14 variants
Breaking the count down: 2 base, 2 instruct and 2 self-speculative make 6; the other 8 are all few-step distilled, covering NFE 2/4/8/16 across 3B and 8B.
The allocation signals a priority: inference speed sits front and center. Eight distilled variants mean users can pick a speed point on the same weights instead of retraining or distilling again.
One self-speculative variant per size reflects a different trade-off: speculative decoding needs a draft model, and shipping many versions would multiply deployment combinations. An official recommended configuration is more useful than a pile that users must tune themselves.
Four cautions for reading these tables
Base and instruct are not comparable. 71.16 is a base six-task average, 70.34 is a distilled nfe8 score, 91.96 is an instruct GSM8K figure. Three numbers, three configurations.
Context is a hard constraint. 4096 base, 8192 for 8B instruct, below what most long-document work requires. No amount of parameter tuning fixes this.
No CPU support, no quantized variants. Deployment is effectively locked to NVIDIA GPUs with bf16 (or framework-side precision conversion). There is no official GGUF or 4-bit path.
Toolchain requirements are recent. Model cards call for PyTorch 2.4+, Transformers 4.55+, and trust_remote_code; supported serving frameworks include vLLM, SGLang and Docker Model Runner. NVIDIA states it will not run its own API and only releases weights.
What "open weights" does and does not mean
Open weights are not open training data and recipe. Sigma publishes weights, an architecture description and a license; the full data composition is not disclosed. For commercial evaluation the license terms matter more than the "open source" label.
Boundaries
- Checkpoint count, tiers, variant classes, license, training volume, context and framework requirements come from the Hugging Face collection and NVIDIA model cards;
- Six-task averages, GSM8K, AIME25, HumanEval, MBPP and the 10-seed note come from model cards and launch materials;
- Throughput and speedups are NVIDIA's own measurements under different hardware conditions, without cross-validation;
- The exact composition of the six-task suite is not itemized in public materials and is not inferred here;
- This article organizes data and flags conventions; it contains no deployment steps or performance promises.
