Atlas Spark Recipe Registry

Official sparkrun recipes for Atlas — the pure-Rust & CUDA LLM inference engine engineered for NVIDIA DGX Spark (GB10 SM121) and prosumer silicon. Recipes target the verified ghcr.io/atlas-inf/atlas-gb10:latest and azeezish/atlas-gb10:latest images and the atlas runtime in sparkrun.

Usage

The @atlas namespace is reserved in sparkrun and pre-registered in standard installs:

# Flagship: Qwen 3.8 Flash-Next (~180B hybrid MoE in ~90 GB resident VRAM)
sparkrun run @atlas/qwen3.8-flash-next-nvfp4 --hosts localhost

# Nemotron 3.5 Lightning 30B with DSpark speculative drafter (sub-10ms decode)
sparkrun run @atlas/nemotron-3.5-lightning-30b-a3b-nvfp4-dspark --hosts localhost

# Qwen 3.8 27B low-latency dense hybrid
sparkrun run @atlas/qwen3.8-27b-nvfp4 --hosts localhost

# Multi-node Expert Parallelism across 2 Sparks (EP=2)
sparkrun run @atlas/deepseek-v4-flash-nvfp4-ep2 --hosts <spark-1>,<spark-2>

If you are on a custom build or older sparkrun release without the default reservation, register it manually:

sparkrun registry add https://github.com/Atlas-Inf/sparkrun-recipes.git

Inspecting and Listing Recipes

The atlas registry ships in sparkrun's catalog. To list or inspect recipes:

sparkrun recipe list --registry atlas                  # List all Atlas recipes
sparkrun recipe show @atlas/qwen3.8-flash-next-nvfp4   # Inspect recipe flags & defaults
sparkrun recipe show @atlas/nemotron-3.5-lightning-30b-a3b-nvfp4-dspark

---

Catalogue (Organized by Vendor & Family)

The recipe catalogue mirrors the structure and prioritization on atlasinference.dev/#models:

1. Qwen

Qwen3.8 (Flagship)

Recipe	Model Checkpoint	Topology	Description
qwen3.8-27b-nvfp4	nvidia/Qwen3.8-27B-NVFP4	Single GB10	DEFAULT FLAGSHIP. 27B dense hybrid GDN + attention, NVFP4, MTP speculative decoding, FP8 KV cache, 23.59 tok/s single-stream
qwen3.8-27b-nvfp4-concurrency	nvidia/Qwen3.8-27B-NVFP4	Single GB10	Concurrency profile measured to beat vLLM from 1 to 128 streams on GB10
qwen3.8-flash-next-nvfp4	nvidia/Qwen3.8-Flash-Next-NVFP4	Single GB10	~180B hybrid MoE on the vendor ModelOpt pack, 32K context / 8K chunked prefill, BF16 KV, MTP K=1, 47.7 GB PLE n-gram parallel pread NVMe offload (util 0.90 — the measured ceiling on the unified 128 GB)
qwen3.8-27b-nvfp4-agentic	nvidia/Qwen3.8-27B-NVFP4	Single GB10	Agentic gate config: thinking ON, BF16 head + BF16 KV, 32K context, MTP K=4, SLAi scheduler
qwen3.8-27b-nvfp4-bfcl	nvidia/Qwen3.8-27B-NVFP4	Single GB10	BFCL v4 agentic tool-use benchmark profile (its bars are tied to this exact config; do not add serve overrides)

Qwen3.6

Recipe	Model Checkpoint	Topology	Description
qwen3.6-35b-a3b-fp8-mtp	Qwen/Qwen3.6-35B-A3B-FP8	Single GB10	Native FP8, BF16 head + BF16 KV, 64K context, MTP K=2, live tool-call streaming
qwen3.6-35b-a3b-nvfp4	nvidia/Qwen3.6-35B-A3B-NVFP4	Single GB10	NVFP4 MoE weights, MTP K=1, calibrated FP8 KV (128K context)
qwen3.6-35b-a3b-fp8-bf16head	Qwen/Qwen3.6-35B-A3B-FP8	Single GB10	32K safe profile with BF16 head/KV
qwen3.6-35b-a3b-fp8-nvfp4head	Qwen/Qwen3.6-35B-A3B-FP8	Single GB10	NVFP4 lm-head sibling, lower memory footprint
qwen3.6-27b-fp8-mtp	Qwen/Qwen3.6-27B-FP8	Single GB10	Dense hybrid SSM+Attn, MTP K=1, 60K context
qwen3.6-27b-nvfp4	nvidia/Qwen3.6-27B-NVFP4	Single GB10	Dense hybrid SSM+Attn, MTP K=1, BF16 KV
qwen3.6-27b-nvfp4-prefill-record	nvidia/Qwen3.6-27B-NVFP4	Single GB10	Prefill-optimized record profile
qwen3.6-27b-nvfp4-unsloth	unsloth/Qwen3.6-27B-NVFP4	Single GB10	Mixed-precision NVFP4/FP8 layout loader, BFCL gate profile
qwen3.6-27b-fp8	Qwen/Qwen3.6-27B-FP8	Single GB10	Pure FP8 dense baseline

Qwen3.5

Recipe	Model Checkpoint	Topology	Description
qwen3.5-35b-a3b-nvfp4	Sehyo/Qwen3.5-35B-A3B-NVFP4	Single GB10	MTP K=2, ~131 tok/s
qwen3.5-27b-dense-nvfp4	Kbenkhaled/Qwen3.5-27B-NVFP4	Single GB10	Dense hybrid GDN + attention
qwen3.5-122b-a10b-nvfp4-single	Sehyo/Qwen3.5-122B-A10B-NVFP4	Single GB10	Tight KV budget, all 256 experts resident on one GB10
qwen3.5-122b-a10b-nvfp4-ep2	Sehyo/Qwen3.5-122B-A10B-NVFP4	EP=2 (2 Sparks)	Dual-node Expert Parallelism + MTP K=2
qwen3.5-0.8b-bf16-atlas	Qwen/Qwen3.5-0.8B	Single GB10	Lightweight edge development target

Qwen Extensions

Recipe	Model Checkpoint	Topology	Description
qwen3-next-80b-a3b-nvfp4	nvidia/Qwen3-Next-80B-A3B-Instruct-NVFP4	Single GB10	SSM + Attention + MoE with MTP speculative decoding
qwen3-coder-next-fp8	Qwen/Qwen3-Coder-Next-FP8	Single GB10	Native FP8 coding target, BF16 KV cache
qwen3-vl-30b-a3b-nvfp4	ig1/Qwen3-VL-30B-A3B-Instruct-NVFP4	Single GB10	Vision-language hybrid MoE with image/video inputs

---

2. Gemma

Subfamily	Recipe	Model Checkpoint	Topology	Description
Gemma-4	gemma-4-26b-a4b-nvfp4	bg-digitalservices/Gemma-4-26B-A4B-it-NVFP4A16	Single GB10	26B / 4B active MoE with GeGLU activation
Gemma-4	gemma-4-31b-nvfp4	nvidia/Gemma-4-31B-IT-NVFP4	Single GB10	Dense 31B, sliding + full attention, gemma4 parser
Gemma Diffusion	diffusion-gemma-bf16	google/diffusion-gemma	Single GB10	BF16 diffusion generative weights
Gemma Diffusion	diffusion-gemma-fp8-dynamic	google/diffusion-gemma	Single GB10	Dynamic FP8 activation profile

---

3. Nemotron

Subfamily	Recipe	Model Checkpoint	Topology	Description
Nemotron-3.5 Lightning	nemotron-3.5-lightning-30b-a3b-nvfp4-dspark	nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4	Single GB10	NEW. Mamba-2 SSM + Attention + MoE with 6-layer DSpark drafter (gamma=4 / K=3 verify), sub-10ms decode
Nemotron-3 Nano	nemotron-3-nano-30b-a3b-nvfp4	nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4	Single GB10	30B / 3B active Mamba-2 hybrid MoE
Nemotron-3 Super	nemotron-3-super-120b-a12b-nvfp4	nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4	Single GB10	120B / 12B active hybrid MoE

---

4. Mistral

Subfamily	Recipe	Model Checkpoint	Topology	Description
Mistral-Small-4	mistral-small-4-119b-nvfp4	mistralai/Mistral-Small-4-119B-2603-NVFP4	Single GB10	MLA attention architecture, mandatory BF16 KV cache

---

5. MiniMax

Subfamily	Recipe	Model Checkpoint	Topology	Description
MiniMax-M2.7	minimax-m2.7-nvfp4-ep2	lukealonso/MiniMax-M2.7-NVFP4	EP=2 (2 Sparks)	229B / ~10B active MoE with 256 experts across 2 nodes

---

6. DeepSeek

Subfamily	Recipe	Model Checkpoint	Topology	Description
DeepSeek-V4	deepseek-v4-flash-nvfp4-ep2	deepseek-ai/DeepSeek-V4-Flash-NVFP4	EP=2 (2 Sparks)	Dual-node Expert Parallelism MoE profile

---

Directory Layout

recipes/
├── qwen3.8/
│   ├── qwen3.8-27b-nvfp4.yaml
│   ├── qwen3.8-27b-nvfp4-agentic.yaml
│   ├── qwen3.8-27b-nvfp4-concurrency.yaml
│   ├── qwen3.8-27b-nvfp4-bfcl.yaml
│   ├── qwen3.8-27b-nvfp4-dflash2.yaml
│   └── qwen3.8-flash-next-nvfp4.yaml
├── qwen3.6/
│   ├── qwen3.6-35b-a3b-fp8-mtp.yaml
│   ├── qwen3.6-35b-a3b-nvfp4.yaml
│   ├── ... (7 additional Qwen3.6 recipes)
├── qwen3.5/
│   ├── qwen3.5-35b-a3b-nvfp4.yaml
│   ├── qwen3.5-27b-dense-nvfp4.yaml
│   ├── qwen3.5-122b-a10b-nvfp4-single.yaml
│   ├── qwen3.5-122b-a10b-nvfp4-ep2.yaml
│   └── qwen3.5-0.8b-bf16-atlas.yaml
├── qwen3-next/
│   └── qwen3-next-80b-a3b-nvfp4.yaml
├── qwen3-coder-next/
│   └── qwen3-coder-next-fp8.yaml
├── qwen3-vl/
│   └── qwen3-vl-30b-a3b-nvfp4.yaml
├── gemma4/
│   ├── gemma-4-26b-a4b-nvfp4.yaml
│   └── gemma-4-31b-nvfp4.yaml
├── diffusion-gemma/
│   ├── diffusion-gemma-bf16.yaml
│   └── diffusion-gemma-fp8-dynamic.yaml
├── nemotron-3.5-lightning/
│   └── nemotron-3.5-lightning-30b-a3b-nvfp4-dspark.yaml
├── nemotron-3-nano/
│   └── nemotron-3-nano-30b-a3b-nvfp4.yaml
├── nemotron-3-super/
│   └── nemotron-3-super-120b-a12b-nvfp4.yaml
├── mistral-small-4/
│   └── mistral-small-4-119b-nvfp4.yaml
├── minimax-m2.7/
│   └── minimax-m2.7-nvfp4-ep2.yaml
└── deepseek-v4/
    └── deepseek-v4-flash-nvfp4-ep2.yaml

Sparkrun recurses through the recipes/ directory structure; each recipe is referenced by qualified stem (e.g. @atlas/qwen3.8-flash-next-nvfp4).

---

Hardware & Runtime Configuration

Key production-validated settings encoded across recipes:

• Qwen 3.8 Flash-Next: Uses nvidia/Qwen3.8-Flash-Next-NVFP4 (the vendor ModelOpt MIXED_PRECISION pack; the earlier RadixArk community mirror shares the same qwen3.8-flash-next kernel target). Binds the specialized SM121 target, BF16 KV cache (avoids clipping without scale tensors), and dynamically streams the 47.7 GB n-gram table from disk to stay under 90 GB VRAM.
• Nemotron 3.5 Lightning: Uses --dflash with drafter nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DSpark. Requires ATLAS_DFLASH_OPTION_B=1. Set ATLAS_NO_TOOL_INJECT=1 for +15.58 BFCL accuracy boost.
• Mistral Small 4: Enforces kv_cache_dtype: bf16 to protect the MLA compressed latent.
• Qwen3-Coder-Next-FP8: Configures ssm_cache_slots: 0, oom_guard_mb: 1024, and kv_cache_dtype: bf16.
• EP=2 Multi-Node: Dual-node configurations (MiniMax M2.7, Qwen3.5-122B, DeepSeek V4) pass matched --speculative and --mtp-quantization across ranks.

---

Related Links

• Website: https://atlasinference.dev
• Atlas Inference Engine: https://github.com/Atlas-Inf/atlas
• Sparkrun: https://github.com/spark-arena/sparkrun
• Docker Image: ghcr.io/atlas-inf/atlas-gb10:latest / azeezish/atlas-gb10:latest
• Discord: Join our Discord
• X: @AtlasInference

---

License

AGPL-3.0 — see LICENSE. Matches the upstream Atlas license.

_{Continuity notice. Atlas is continuing. This repository, the Atlas-Inf GitHub organization, and atlasinference.dev are the replacement official Atlas channels. The existing website and GitHub repository remain disputed Atlas assets that have not been relinquished.}