mlx-stack 0.1.0__py3-none-any.whl

mlx_stack/data/__init__.py

@@ -0,0 +1 @@
+"""Static data files for mlx-stack."""

mlx_stack/data/catalog/__init__.py

@@ -0,0 +1 @@
+"""Model catalog YAML data files."""

mlx_stack/data/catalog/deepseek-r1-32b.yaml

@@ -0,0 +1,46 @@
+id: deepseek-r1-32b
+name: DeepSeek R1 32B
+family: DeepSeek R1
+params_b: 32.0
+architecture: mamba2-hybrid
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/DeepSeek-R1-0528-Qwen3-32B-4bit
+    disk_size_gb: 18.0
+  int8:
+    hf_repo: mlx-community/DeepSeek-R1-0528-Qwen3-32B-8bit
+    disk_size_gb: 34.0
+  bf16:
+    hf_repo: deepseek-ai/DeepSeek-R1-0528-Qwen3-32B
+    disk_size_gb: 65.0
+    convert_from: true
+capabilities:
+  tool_calling: false
+  tool_call_parser: null
+  thinking: true
+  reasoning_parser: deepseek_r1
+  vision: false
+quality:
+  overall: 85
+  coding: 86
+  reasoning: 90
+  instruction_following: 80
+benchmarks:
+  m4-pro-48:
+    prompt_tps: 26.0
+    gen_tps: 15.0
+    memory_gb: 20.0
+  m4-max-128:
+    prompt_tps: 40.0
+    gen_tps: 23.0
+    memory_gb: 20.0
+  m5-max-128:
+    prompt_tps: 44.0
+    gen_tps: 25.0
+    memory_gb: 20.0
+tags:
+  - reasoning
+  - thinking
+  - quality
+  - long-context

mlx_stack/data/catalog/deepseek-r1-8b.yaml

@@ -0,0 +1,45 @@
+id: deepseek-r1-8b
+name: DeepSeek R1 8B
+family: DeepSeek R1
+params_b: 8.0
+architecture: mamba2-hybrid
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/DeepSeek-R1-0528-Qwen3-8B-4bit
+    disk_size_gb: 4.5
+  int8:
+    hf_repo: mlx-community/DeepSeek-R1-0528-Qwen3-8B-8bit
+    disk_size_gb: 8.5
+  bf16:
+    hf_repo: deepseek-ai/DeepSeek-R1-0528-Qwen3-8B
+    disk_size_gb: 16.0
+    convert_from: true
+capabilities:
+  tool_calling: false
+  tool_call_parser: null
+  thinking: true
+  reasoning_parser: deepseek_r1
+  vision: false
+quality:
+  overall: 70
+  coding: 72
+  reasoning: 75
+  instruction_following: 64
+benchmarks:
+  m4-pro-48:
+    prompt_tps: 90.0
+    gen_tps: 50.0
+    memory_gb: 5.5
+  m4-max-128:
+    prompt_tps: 135.0
+    gen_tps: 74.0
+    memory_gb: 5.5
+  m5-max-128:
+    prompt_tps: 148.0
+    gen_tps: 82.0
+    memory_gb: 5.5
+tags:
+  - reasoning
+  - thinking
+  - long-context

mlx_stack/data/catalog/gemma3-12b.yaml

@@ -0,0 +1,45 @@
+id: gemma3-12b
+name: Gemma 3 12B
+family: Gemma 3
+params_b: 12.0
+architecture: transformer
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/gemma-3-12b-it-4bit
+    disk_size_gb: 7.0
+  int8:
+    hf_repo: mlx-community/gemma-3-12b-it-8bit
+    disk_size_gb: 13.0
+  bf16:
+    hf_repo: google/gemma-3-12b-it
+    disk_size_gb: 25.0
+    convert_from: true
+capabilities:
+  tool_calling: false
+  tool_call_parser: null
+  thinking: false
+  reasoning_parser: null
+  vision: true
+quality:
+  overall: 72
+  coding: 68
+  reasoning: 70
+  instruction_following: 75
+benchmarks:
+  m4-pro-48:
+    prompt_tps: 62.0
+    gen_tps: 35.0
+    memory_gb: 8.2
+  m4-max-128:
+    prompt_tps: 90.0
+    gen_tps: 52.0
+    memory_gb: 8.2
+  m5-max-128:
+    prompt_tps: 100.0
+    gen_tps: 57.0
+    memory_gb: 8.2
+tags:
+  - vision
+  - balanced
+  - multimodal

mlx_stack/data/catalog/gemma3-27b.yaml

@@ -0,0 +1,45 @@
+id: gemma3-27b
+name: Gemma 3 27B
+family: Gemma 3
+params_b: 27.0
+architecture: transformer
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/gemma-3-27b-it-4bit
+    disk_size_gb: 15.0
+  int8:
+    hf_repo: mlx-community/gemma-3-27b-it-8bit
+    disk_size_gb: 28.0
+  bf16:
+    hf_repo: google/gemma-3-27b-it
+    disk_size_gb: 54.0
+    convert_from: true
+capabilities:
+  tool_calling: false
+  tool_call_parser: null
+  thinking: false
+  reasoning_parser: null
+  vision: true
+quality:
+  overall: 80
+  coding: 76
+  reasoning: 78
+  instruction_following: 83
+benchmarks:
+  m4-pro-48:
+    prompt_tps: 30.0
+    gen_tps: 18.0
+    memory_gb: 17.0
+  m4-max-128:
+    prompt_tps: 45.0
+    gen_tps: 26.0
+    memory_gb: 17.0
+  m5-max-128:
+    prompt_tps: 50.0
+    gen_tps: 29.0
+    memory_gb: 17.0
+tags:
+  - vision
+  - quality
+  - multimodal

mlx_stack/data/catalog/gemma3-4b.yaml

@@ -0,0 +1,45 @@
+id: gemma3-4b
+name: Gemma 3 4B
+family: Gemma 3
+params_b: 4.0
+architecture: transformer
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/gemma-3-4b-it-4bit
+    disk_size_gb: 2.3
+  int8:
+    hf_repo: mlx-community/gemma-3-4b-it-8bit
+    disk_size_gb: 4.4
+  bf16:
+    hf_repo: google/gemma-3-4b-it
+    disk_size_gb: 8.5
+    convert_from: true
+capabilities:
+  tool_calling: false
+  tool_call_parser: null
+  thinking: false
+  reasoning_parser: null
+  vision: true
+quality:
+  overall: 52
+  coding: 45
+  reasoning: 48
+  instruction_following: 58
+benchmarks:
+  m4-pro-48:
+    prompt_tps: 160.0
+    gen_tps: 78.0
+    memory_gb: 3.2
+  m4-max-128:
+    prompt_tps: 230.0
+    gen_tps: 115.0
+    memory_gb: 3.2
+  m5-max-128:
+    prompt_tps: 250.0
+    gen_tps: 125.0
+    memory_gb: 3.2
+tags:
+  - vision
+  - fast-inference
+  - multimodal

mlx_stack/data/catalog/llama3.3-8b.yaml

@@ -0,0 +1,44 @@
+id: llama3.3-8b
+name: Llama 3.3 8B
+family: Llama 3.3
+params_b: 8.0
+architecture: transformer
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/Llama-3.3-8B-Instruct-4bit
+    disk_size_gb: 4.5
+  int8:
+    hf_repo: mlx-community/Llama-3.3-8B-Instruct-8bit
+    disk_size_gb: 8.5
+  bf16:
+    hf_repo: meta-llama/Llama-3.3-8B-Instruct
+    disk_size_gb: 16.0
+    convert_from: true
+capabilities:
+  tool_calling: true
+  tool_call_parser: llama3_json
+  thinking: false
+  reasoning_parser: null
+  vision: false
+quality:
+  overall: 63
+  coding: 60
+  reasoning: 58
+  instruction_following: 67
+benchmarks:
+  m4-pro-48:
+    prompt_tps: 100.0
+    gen_tps: 55.0
+    memory_gb: 5.5
+  m4-max-128:
+    prompt_tps: 148.0
+    gen_tps: 82.0
+    memory_gb: 5.5
+  m5-max-128:
+    prompt_tps: 162.0
+    gen_tps: 90.0
+    memory_gb: 5.5
+tags:
+  - balanced
+  - agent-ready

mlx_stack/data/catalog/nemotron-49b.yaml

@@ -0,0 +1,41 @@
+id: nemotron-49b
+name: Nemotron 49B
+family: Nemotron
+params_b: 49.0
+architecture: transformer
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/Nemotron-Ultra-49B-4bit
+    disk_size_gb: 27.0
+  int8:
+    hf_repo: mlx-community/Nemotron-Ultra-49B-8bit
+    disk_size_gb: 52.0
+  bf16:
+    hf_repo: nvidia/Nemotron-Ultra-49B
+    disk_size_gb: 100.0
+    convert_from: true
+capabilities:
+  tool_calling: true
+  tool_call_parser: hermes
+  thinking: true
+  reasoning_parser: nemotron
+  vision: false
+quality:
+  overall: 87
+  coding: 85
+  reasoning: 88
+  instruction_following: 88
+benchmarks:
+  m4-max-128:
+    prompt_tps: 22.0
+    gen_tps: 13.0
+    memory_gb: 30.0
+  m5-max-128:
+    prompt_tps: 25.0
+    gen_tps: 15.0
+    memory_gb: 30.0
+tags:
+  - quality
+  - thinking
+  - agent-ready

mlx_stack/data/catalog/nemotron-8b.yaml

@@ -0,0 +1,44 @@
+id: nemotron-8b
+name: Nemotron 8B
+family: Nemotron
+params_b: 8.0
+architecture: transformer
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/Nemotron-Mini-8B-4bit
+    disk_size_gb: 4.5
+  int8:
+    hf_repo: mlx-community/Nemotron-Mini-8B-8bit
+    disk_size_gb: 8.5
+  bf16:
+    hf_repo: nvidia/Nemotron-Mini-8B
+    disk_size_gb: 16.0
+    convert_from: true
+capabilities:
+  tool_calling: true
+  tool_call_parser: hermes
+  thinking: false
+  reasoning_parser: null
+  vision: false
+quality:
+  overall: 65
+  coding: 60
+  reasoning: 58
+  instruction_following: 70
+benchmarks:
+  m4-pro-48:
+    prompt_tps: 98.0
+    gen_tps: 54.0
+    memory_gb: 5.5
+  m4-max-128:
+    prompt_tps: 145.0
+    gen_tps: 80.0
+    memory_gb: 5.5
+  m5-max-128:
+    prompt_tps: 160.0
+    gen_tps: 88.0
+    memory_gb: 5.5
+tags:
+  - balanced
+  - agent-ready

mlx_stack/data/catalog/qwen3-8b.yaml

@@ -0,0 +1,45 @@
+id: qwen3-8b
+name: Qwen 3 8B
+family: Qwen 3
+params_b: 8.0
+architecture: transformer
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/Qwen3-8B-4bit
+    disk_size_gb: 4.5
+  int8:
+    hf_repo: mlx-community/Qwen3-8B-8bit
+    disk_size_gb: 8.5
+  bf16:
+    hf_repo: Qwen/Qwen3-8B
+    disk_size_gb: 16.0
+    convert_from: true
+capabilities:
+  tool_calling: true
+  tool_call_parser: hermes
+  thinking: true
+  reasoning_parser: qwen3
+  vision: false
+quality:
+  overall: 64
+  coding: 62
+  reasoning: 60
+  instruction_following: 68
+benchmarks:
+  m4-pro-48:
+    prompt_tps: 96.0
+    gen_tps: 53.0
+    memory_gb: 5.5
+  m4-max-128:
+    prompt_tps: 142.0
+    gen_tps: 78.0
+    memory_gb: 5.5
+  m5-max-128:
+    prompt_tps: 156.0
+    gen_tps: 86.0
+    memory_gb: 5.5
+tags:
+  - balanced
+  - agent-ready
+  - thinking

mlx_stack/data/catalog/qwen3.5-0.8b.yaml

@@ -0,0 +1,45 @@
+id: qwen3.5-0.8b
+name: Qwen 3.5 0.8B
+family: Qwen 3.5
+params_b: 0.8
+architecture: transformer
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/Qwen3.5-0.8B-4bit
+    disk_size_gb: 0.65
+  int8:
+    hf_repo: mlx-community/Qwen3.5-0.8B-8bit
+    disk_size_gb: 0.9
+  bf16:
+    hf_repo: Qwen/Qwen3.5-0.8B
+    disk_size_gb: 1.6
+    convert_from: true
+capabilities:
+  tool_calling: true
+  tool_call_parser: hermes
+  thinking: false
+  reasoning_parser: null
+  vision: false
+quality:
+  overall: 42
+  coding: 38
+  reasoning: 35
+  instruction_following: 48
+benchmarks:
+  m4-pro-48:
+    prompt_tps: 320.0
+    gen_tps: 125.0
+    memory_gb: 0.8
+  m4-max-128:
+    prompt_tps: 480.0
+    gen_tps: 185.0
+    memory_gb: 0.8
+  m5-max-128:
+    prompt_tps: 510.0
+    gen_tps: 195.0
+    memory_gb: 0.8
+tags:
+  - lightweight
+  - fast-inference
+  - agent-ready

mlx_stack/data/catalog/qwen3.5-14b.yaml

@@ -0,0 +1,46 @@
+id: qwen3.5-14b
+name: Qwen 3.5 14B
+family: Qwen 3.5
+params_b: 14.0
+architecture: transformer
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/Qwen3.5-14B-4bit
+    disk_size_gb: 8.0
+  int8:
+    hf_repo: mlx-community/Qwen3.5-14B-8bit
+    disk_size_gb: 15.0
+  bf16:
+    hf_repo: Qwen/Qwen3.5-14B
+    disk_size_gb: 28.0
+    convert_from: true
+capabilities:
+  tool_calling: true
+  tool_call_parser: hermes
+  thinking: true
+  reasoning_parser: qwen3
+  vision: false
+quality:
+  overall: 76
+  coding: 74
+  reasoning: 72
+  instruction_following: 80
+benchmarks:
+  m4-pro-48:
+    prompt_tps: 55.0
+    gen_tps: 32.0
+    memory_gb: 9.5
+  m4-max-128:
+    prompt_tps: 82.0
+    gen_tps: 48.0
+    memory_gb: 9.5
+  m5-max-128:
+    prompt_tps: 90.0
+    gen_tps: 53.0
+    memory_gb: 9.5
+tags:
+  - balanced
+  - agent-ready
+  - thinking
+  - quality

mlx_stack/data/catalog/qwen3.5-32b.yaml

@@ -0,0 +1,45 @@
+id: qwen3.5-32b
+name: Qwen 3.5 32B
+family: Qwen 3.5
+params_b: 32.0
+architecture: transformer
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/Qwen3.5-32B-4bit
+    disk_size_gb: 18.0
+  int8:
+    hf_repo: mlx-community/Qwen3.5-32B-8bit
+    disk_size_gb: 34.0
+  bf16:
+    hf_repo: Qwen/Qwen3.5-32B
+    disk_size_gb: 65.0
+    convert_from: true
+capabilities:
+  tool_calling: true
+  tool_call_parser: hermes
+  thinking: true
+  reasoning_parser: qwen3
+  vision: false
+quality:
+  overall: 84
+  coding: 82
+  reasoning: 83
+  instruction_following: 86
+benchmarks:
+  m4-pro-48:
+    prompt_tps: 28.0
+    gen_tps: 16.0
+    memory_gb: 20.0
+  m4-max-128:
+    prompt_tps: 42.0
+    gen_tps: 24.0
+    memory_gb: 20.0
+  m5-max-128:
+    prompt_tps: 46.0
+    gen_tps: 27.0
+    memory_gb: 20.0
+tags:
+  - quality
+  - agent-ready
+  - thinking

mlx_stack/data/catalog/qwen3.5-3b.yaml

@@ -0,0 +1,44 @@
+id: qwen3.5-3b
+name: Qwen 3.5 3B
+family: Qwen 3.5
+params_b: 3.0
+architecture: transformer
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/Qwen3.5-3B-4bit
+    disk_size_gb: 1.8
+  int8:
+    hf_repo: mlx-community/Qwen3.5-3B-8bit
+    disk_size_gb: 3.4
+  bf16:
+    hf_repo: Qwen/Qwen3.5-3B
+    disk_size_gb: 6.5
+    convert_from: true
+capabilities:
+  tool_calling: true
+  tool_call_parser: hermes
+  thinking: false
+  reasoning_parser: null
+  vision: false
+quality:
+  overall: 55
+  coding: 50
+  reasoning: 48
+  instruction_following: 60
+benchmarks:
+  m4-pro-48:
+    prompt_tps: 180.0
+    gen_tps: 88.0
+    memory_gb: 2.5
+  m4-max-128:
+    prompt_tps: 260.0
+    gen_tps: 130.0
+    memory_gb: 2.5
+  m5-max-128:
+    prompt_tps: 280.0
+    gen_tps: 140.0
+    memory_gb: 2.5
+tags:
+  - fast-inference
+  - agent-ready

mlx_stack/data/catalog/qwen3.5-72b.yaml

@@ -0,0 +1,42 @@
+id: qwen3.5-72b
+name: Qwen 3.5 72B
+family: Qwen 3.5
+params_b: 72.0
+architecture: transformer
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/Qwen3.5-72B-4bit
+    disk_size_gb: 40.0
+  int8:
+    hf_repo: mlx-community/Qwen3.5-72B-8bit
+    disk_size_gb: 75.0
+  bf16:
+    hf_repo: Qwen/Qwen3.5-72B
+    disk_size_gb: 145.0
+    convert_from: true
+capabilities:
+  tool_calling: true
+  tool_call_parser: hermes
+  thinking: true
+  reasoning_parser: qwen3
+  vision: false
+quality:
+  overall: 91
+  coding: 90
+  reasoning: 92
+  instruction_following: 92
+benchmarks:
+  m4-max-128:
+    prompt_tps: 18.0
+    gen_tps: 11.0
+    memory_gb: 42.0
+  m5-max-128:
+    prompt_tps: 20.0
+    gen_tps: 12.5
+    memory_gb: 42.0
+tags:
+  - premium
+  - quality
+  - thinking
+  - agent-ready

mlx_stack/data/catalog/qwen3.5-8b.yaml

@@ -0,0 +1,45 @@
+id: qwen3.5-8b
+name: Qwen 3.5 8B
+family: Qwen 3.5
+params_b: 8.0
+architecture: transformer
+min_mlx_lm_version: "0.22.0"
+sources:
+  int4:
+    hf_repo: mlx-community/Qwen3.5-8B-4bit
+    disk_size_gb: 4.5
+  int8:
+    hf_repo: mlx-community/Qwen3.5-8B-8bit
+    disk_size_gb: 8.5
+  bf16:
+    hf_repo: Qwen/Qwen3.5-8B
+    disk_size_gb: 16.0
+    convert_from: true
+capabilities:
+  tool_calling: true
+  tool_call_parser: hermes
+  thinking: true
+  reasoning_parser: qwen3
+  vision: false
+quality:
+  overall: 68
+  coding: 65
+  reasoning: 62
+  instruction_following: 72
+benchmarks:
+  m4-pro-48:
+    prompt_tps: 95.0
+    gen_tps: 52.0
+    memory_gb: 5.5
+  m4-max-128:
+    prompt_tps: 140.0
+    gen_tps: 77.0
+    memory_gb: 5.5
+  m5-max-128:
+    prompt_tps: 155.0
+    gen_tps: 85.0
+    memory_gb: 5.5
+tags:
+  - balanced
+  - agent-ready
+  - thinking

mlx_stack/py.typed

@@ -0,0 +1 @@
+

mlx_stack/utils/__init__.py

@@ -0,0 +1 @@
+"""Utility modules for mlx-stack."""