Mainline llama.cpp cannot convert or run this architecture. Use spiritbuun/buun-llama-cpp — its converter maps NemotronLabsDiffusionModel to the GGUF dream architecture.

Files

File	Size
`Nemotron-Labs-Diffusion-8B-Q4_K_M.gguf`	4.84 GB
`Nemotron-Labs-Diffusion-8B-Q5_K_M.gguf`	5.64 GB

Run

Diffusion (dlm):

bash

llama-diffusion-cli -m Nemotron-Labs-Diffusion-8B-Q4_K_M.gguf -ngl 99 -fa on -p "your prompt" \
  --diffusion-block-length 32 --diffusion-steps 128 --diffusion-threshold 0.9 -ub 512

Self-speculation:

bash

llama-diffusion-cli -m Nemotron-Labs-Diffusion-8B-Q4_K_M.gguf -ngl 99 -fa on -p "your prompt" \
  --diffusion-self-spec --diffusion-block-length 32 --diffusion-draft-length 16 --diffusion-steps 128

OpenAI-compatible server (auto-enables self-speculation = AR-equivalent):

llama-server -m Nemotron-Labs-Diffusion-8B-Q4_K_M.gguf -ngl 99 -c 4096 -fa on --port 8080

Benchmarks (single 16 GB-VRAM GPU, Q4_K_M unless noted)

Model	Engine / mode	tok/s	TTFT	First-visible	VRAM (weights)	GPU total	Streaming	Quality
gemma-4-E4B	Ollama (AR)	148.7	0.03s	~0.05s	3.11 GB	6.28 GB	◎ per-token	good
gemma-4-12B	Ollama (AR)	84.4	0.04s	~0.06s	7.27 GB	9.69 GB	◎ per-token	good (clean numbered list)
Nemotron-8B Q4	buun dlm	69.4	—	~0.53s	4.84 GB	~7.0 GB	△ block-wise	good (slightly rough)
Nemotron-8B Q5	buun dlm	58.7	—	~0.56s	5.64 GB	~7.5 GB	△ block-wise	good
Nemotron-8B Q4	llama-server	~46	prefill 454/s	streams*	4.84 GB	~7.0 GB	→ reasoning_content	good

First-visible = time until the first text can actually be shown (warm). AR streams from the first token (≈ TTFT); block diffusion must finish the first 32-token block before anything is displayable, so its first-visible (~0.5 s) is much higher than its TTFT suggests.
*The server (self-spec) streams per accepted chunk, but the answer is gated behind reasoning_content.
gemma-4 rows are public autoregressive GGUFs measured on the same GPU for reference.

Text-quality comparison

Same prompt, chat-templated, temp 0.7 — "犬と猫の違いを3つ、箇条書きで簡潔に教えてください。" (3 bullet points contrasting dogs vs cats). Nemotron via llama-server (answer in reasoning_content).

Model	Output excerpt	Follows format / contrast
gemma-4-12B	群れの性質：犬は「群れ」で行動し飼い主を家族として従う。一方、猫は「単独」で自立心が強い…	✓ clean 3-bullet contrast
gemma-4-E4B	犬は群れ（家族）を意識し指示に従う傾向。一方、猫は単独行動を好み自立した性質…	✓ clean
Nemotron-8B Q5	犬は社会的で家族との絆が強い。犬は遊びに応じ行動を変える…（犬中心で対比が弱い）	△ weaker contrast
Nemotron-14B Q5	食べ方：犬は肉食動物として肉を食べる。猫も肉食だが自分で捕まえて食べる…（やや冗長）	◯ but slightly circular

Empathy prompt ("最近ちょっと疲れてるんだ。やさしく励まして。"): all four give warm, fluent replies.

Takeaway: all produce fluent Japanese. The optimized autoregressive gemma-4 models follow structured instructions most cleanly; the Nemotron diffusion models are coherent but a bit rougher (more repetition / weaker factual contrast). Small single-prompt sample — treat as indicative, not a leaderboard.

Modes / caveats

diffusion (dlm) and self-speculation both run via llama-diffusion-cli.
llama-server auto-detects the model and runs self-speculation (diffusion draft + AR verify, lossless vs. AR at temperature 0). This is the AR path for GGUF.
Self-speculation gives no speedup over plain block diffusion on a single 16 GB consumer GPU (its benefit appears on large datacenter GPUs).
The generic llama-cli does not work with this dream arch (it hangs) — use llama-server or llama-diffusion-cli.
llama-server emits the answer in reasoning_content; route it to content with the appropriate --reasoning option.
Optimized AR GGUFs (e.g. gemma-4) are faster and stream per-token; this model emits 32-token blocks.

License

Derivative of NVIDIA's model — governed by the NVIDIA Open Model License.

Credits

Original model: NVIDIA — nvidia/Nemotron-Labs-Diffusion-8B
Diffusion-aware llama.cpp fork & 14B GGUF: spiritbuun

suzukimain/Nemotron-Labs-Diffusion-8B-GGUF

Nemotron-Labs-Diffusion-8B — GGUF

⚠️ Requires a diffusion-aware llama.cpp fork