Single DGX Spark Recipes, optimized for single spark to single user workloads. All of these recipes can be executed with the `sparkrun run <filename>` command.

Sparkrun-MiniMax-M2.5-230B-MXFP4.yaml

metadata:
  description: |
    MiniMax is the largest model physically capable of fitting on a single Spark.
    Note: It does run slower than Qwen due to lack of MTP.
    Achieves ~18 tk/s.
  model_dtype: MXFP4
  num_kv_heads: 2
  head_dim: 256
  num_layers: 48
  model_params: 122B

name: exdysa/MiniMax-M2.5-REAP-139B-A10B-GGUF-MXFP4_MOE
runtime: llama-cpp
model: exdysa/MiniMax-M2.5-REAP-139B-A10B-GGUF-MXFP4_MOE
recipe_version: '1'
solo_only: true
max_nodes: 1
container: scitrera/dgx-spark-llama-cpp:latest

command: |
  llama-server \
    -m {model} \
    --host {host} \
    --port {port} \
    --n-gpu-layers {n_gpu_layers} \
    --batch-size {batch_size} \
    --ubatch-size {ubatch_size} \
    --cache-type-k {cache_type_k} \
    --cache-type-v {cache_type_v} \
    --no-mmap \
    --keep {keep} \
    --cache-prompt \
    --cache-reuse {cache_reuse} \
    --parallel {parallel} \
    --spec-type {spec_type} \
    --no-ui \
    --threads {threads} \
    --flash-attn {flash_attn} \
    --ctx-size {context_size} \
    --ctx-checkpoints {context_checkpoints} \
    --cache-ram {cache_ram}

defaults:
  host: '0.0.0.0'
  port: 8000
  n_gpu_layers: 99
  batch_size: 4096
  ubatch_size: 2048
  cache_type_k: 'q8_0'
  cache_type_v: 'q8_0'
  cache_ram: 1024
  cache_reuse: 128
  context_size: 196608
  context_checkpoints: 10
  flash_attn: 'on'
  keep: -1
  parallel: 1
  threads: 1
  spec_type: 'ngram-cache'

env:
  TZ: America/Los_Angeles

Sparkrun-Qwen3.5-122B-A10B-MTP-MXFP4.yaml

metadata:
  description: |
    Maximized speed variant of Qwen3.5 122B for 1 user -> 1 spark.
    Strict focus on TTFT and token generation speed, at the cost
    of some intelligence.
    Achieves ~40 tk/s.
  model_dtype: MXFP4
  num_kv_heads: 2
  head_dim: 256
  num_layers: 48
  model_params: 122B
 
name: unsloth/Qwen3.5-122B-A10B-MTP-GGUF-MXFP4_MOE
runtime: llama-cpp
model: unsloth/Qwen3.5-122B-A10B-MTP-GGUF:MXFP4_MOE
recipe_version: '1'
solo_only: true
max_nodes: 1
container: scitrera/dgx-spark-llama-cpp:latest

  llama-server \
    -m {model} \
    --host {host} \
    --port {port} \
    --n-gpu-layers {n_gpu_layers} \
    --batch-size {batch_size} \
    --ubatch-size {ubatch_size} \
    --cache-type-k {cache_type_k} \
    --cache-type-v {cache_type_v} \
    --no-mmap \
    --keep {keep} \
    --cache-prompt \
    --cache-reuse {cache_reuse} \
    --parallel {parallel} \
    --spec-type {spec_type} \
    --spec-draft-n-max {spec_draft_n_max} \
    --spec-draft-n-min {spec_draft_n_min} \
    --no-ui \
    --threads {threads} \
    --flash-attn {flash_attn} \
    --ctx-size {context_size} \
    --ctx-checkpoints {context_checkpoints} \
    --cache-ram {cache_ram}

defaults:
  host: '0.0.0.0'
  port: 8000
  n_gpu_layers: 99
  batch_size: 4096
  ubatch_size: 2048
  cache_type_k: 'f16'
  cache_type_v: 'f16'
  cache_ram: 2048
  cache_reuse: 128
  context_size: 262144
  context_checkpoints: 24
  flash_attn: 'on'
  keep: -1
  parallel: 1
  threads: 1
  spec_type: 'draft-mtp'
  spec_draft_n_max: 3
  spec_draft_n_min: 0.6
  
env:
  TZ: America/Los_Angeles

Sparkrun-Qwen3.5-122B-A10B-MTP-Q5-K-M.yaml

metadata:
  description: |
    Maximized intelligence variant of Qwen3.5 122B for 1 user -> 1 spark.
    Strict focus on using the least lobotomized quant possible that will
    fit on a spark, alongside Qwen's full context size. While it is possible
    to fit smarter quants on a Spark, you would need to sacrifice context size.
    A strong attempt was made to retain TTFT and token generation speed via
    optimized flags.
    This recipe pushes the DGX Spark platform to the limits, and just *barely*
    does not cause OOM. My spark's idle RAM usage is 3 GB with no models loaded.
    If your idle RAM usage is higher than mine, you might hit OOM.
    Achieves ~32 tk/s.
  model_dtype: Q5_K_M
  num_kv_heads: 2
  head_dim: 256
  num_layers: 48
  model_params: 122B

name: unsloth/Qwen3.5-122B-A10B-MTP-GGUF-Q5_K_M
runtime: llama-cpp
model: unsloth/Qwen3.5-122B-A10B-MTP-GGUF:Q5_K_M
recipe_version: '1'
solo_only: true
max_nodes: 1
container: scitrera/dgx-spark-llama-cpp:latest

command: |
  llama-server \
    -m {model} \
    --host {host} \
    --port {port} \
    --n-gpu-layers {n_gpu_layers} \
    --batch-size {batch_size} \
    --ubatch-size {ubatch_size} \
    --cache-type-k {cache_type_k} \
    --cache-type-v {cache_type_v} \
    --no-mmap \
    --keep {keep} \
    --cache-prompt \
    --cache-reuse {cache_reuse} \
    --parallel {parallel} \
    --spec-type {spec_type} \
    --spec-draft-n-max {spec_draft_n_max} \
    --spec-draft-n-min {spec_draft_n_min} \
    --no-wui \
    --threads {threads} \
    --flash-attn {flash_attn} \
    --ctx-size {context_size} \
    --ctx-checkpoints {context_checkpoints} \
    --cache-ram {cache_ram}

defaults:
  host: '0.0.0.0'
  port: 8000
  n_gpu_layers: 99
  batch_size: 4096
  ubatch_size: 2048
  cache_type_k: 'q8_0'
  cache_type_v: 'q8_0'
  cache_ram: 1024
  cache_reuse: 128
  context_size: 262144
  context_checkpoints: 12
  flash_attn: 'on'
  keep: -1
  parallel: 1
  threads: 1
  spec_type: 'draft-mtp'
  spec_draft_n_max: 3
  spec_draft_n_min: 0.6

env:
  TZ: America/Los_Angeles