# Farnsworth Model Configurations
# Updated January 2025 with latest efficient models
models:
# ============================================
# TIER 1: Ultra-Efficient (< 2GB VRAM/RAM)
# Best for: Edge devices, CPU-only, speed
# ============================================
qwen3-0.6b:
name: "Qwen3-0.6B"
description: "Ultra-lightweight with 100+ language support"
backend: "ollama"
ollama_name: "qwen3:0.6b"
llama_cpp_repo: "Qwen/Qwen3-0.6B-GGUF"
llama_cpp_file: "qwen3-0.6b-q4_k_m.gguf"
params: 0.6e9
vram_gb: 1.0
ram_gb: 2.0
quantization: "Q4_K_M"
context_length: 32768
strengths:
- "multilingual"
- "lightweight"
- "fast"
- "edge"
license: "Apache-2.0"
recommended_for:
- "lightweight"
- "multilingual"
- "edge"
release_date: "2025-05"
tinyllama-1.1b:
name: "TinyLlama-1.1B"
description: "Fastest inference, optimized for edge devices"
backend: "ollama"
ollama_name: "tinyllama:1.1b"
llama_cpp_repo: "TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF"
llama_cpp_file: "tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf"
params: 1.1e9
vram_gb: 0.8
ram_gb: 2.0
quantization: "Q4_K_M"
context_length: 2048
strengths:
- "speed"
- "edge"
- "mobile"
- "fast"
license: "Apache-2.0"
recommended_for:
- "speed"
- "edge"
- "mobile"
release_date: "2024-01"
bitnet-2b:
name: "Microsoft BitNet b1.58-2B-4T"
description: "Native 1-bit quantization, 5-7x faster on CPU"
backend: "bitnet"
bitnet_repo: "microsoft/BitNet-b1.58-2B-4T"
params: 2.0e9
vram_gb: 0.5
ram_gb: 1.0
quantization: "1-bit"
context_length: 4096
strengths:
- "speed"
- "efficiency"
- "cpu-optimized"
- "low-power"
license: "MIT"
recommended_for:
- "speed"
- "cpu-only"
- "low-power"
release_date: "2024-10"
# ============================================
# TIER 2: Compact Quality (2-4GB VRAM/RAM)
# Best for: General use, good quality/speed balance
# ============================================
deepseek-r1-1.5b:
name: "DeepSeek-R1-Distill-Qwen-1.5B"
description: "o1-style reasoning in 1.5B params - best small reasoning model"
backend: "ollama"
ollama_name: "deepseek-r1:1.5b"
llama_cpp_repo: "unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF"
llama_cpp_file: "DeepSeek-R1-Distill-Qwen-1.5B-Q4_K_M.gguf"
params: 1.5e9
vram_gb: 2.0
ram_gb: 4.0
quantization: "Q4_K_M"
context_length: 32768
strengths:
- "reasoning"
- "math"
- "code"
- "chain-of-thought"
license: "MIT"
recommended_for:
- "default"
- "reasoning"
- "code"
release_date: "2025-01"
smollm2-1.7b:
name: "SmolLM2-1.7B"
description: "State-of-the-art compact model, beats Qwen2.5-1.5B"
backend: "ollama"
ollama_name: "smollm2:1.7b"
llama_cpp_repo: "HuggingFaceTB/SmolLM2-1.7B-Instruct-GGUF"
llama_cpp_file: "smollm2-1.7b-instruct-q4_k_m.gguf"
params: 1.7e9
vram_gb: 1.5
ram_gb: 3.0
quantization: "Q4_K_M"
context_length: 8192
strengths:
- "instruction-following"
- "quality"
- "on-device"
license: "Apache-2.0"
recommended_for:
- "quality"
- "on-device"
release_date: "2024-11"
gemma-1b:
name: "Gemma-1B"
description: "Google's efficient small model"
backend: "ollama"
ollama_name: "gemma:1b"
llama_cpp_repo: "google/gemma-1.1-2b-it-GGUF"
llama_cpp_file: "gemma-1.1-2b-it.Q4_K_M.gguf"
params: 1.0e9
vram_gb: 1.0
ram_gb: 2.0
quantization: "Q4_K_M"
context_length: 8192
strengths:
- "instruction-following"
- "efficient"
license: "Gemma"
recommended_for:
- "general"
- "efficient"
release_date: "2024-02"
qwen3-4b:
name: "Qwen3-4B"
description: "Best for 4GB budget - MMLU-Pro 74%, excellent reasoning"
backend: "ollama"
ollama_name: "qwen3:4b"
llama_cpp_repo: "Qwen/Qwen3-4B-GGUF"
llama_cpp_file: "qwen3-4b-q4_k_m.gguf"
params: 4.0e9
vram_gb: 2.75
ram_gb: 5.0
quantization: "Q4_K_M"
context_length: 32768
strengths:
- "reasoning"
- "multilingual"
- "code"
- "math"
license: "Apache-2.0"
recommended_for:
- "balanced"
- "4gb-budget"
release_date: "2025-07"
# ============================================
# TIER 3: High Quality (4-8GB VRAM/RAM)
# Best for: Complex tasks, best quality
# ============================================
phi-4-mini:
name: "Phi-4-mini"
description: "GPT-3.5 class, beats 7B/8B in math - 128K context"
backend: "ollama"
ollama_name: "phi4:mini"
llama_cpp_repo: "microsoft/Phi-4-mini-instruct-gguf"
llama_cpp_file: "Phi-4-mini-instruct-Q4_K_M.gguf"
params: 3.8e9
vram_gb: 3.0
ram_gb: 6.0
quantization: "Q4_K_M"
context_length: 128000
strengths:
- "reasoning"
- "math"
- "code"
- "instruction-following"
- "long-context"
license: "MIT"
recommended_for:
- "quality"
- "reasoning"
- "math"
release_date: "2025-05"
phi-4-mini-reasoning:
name: "Phi-4-mini-reasoning"
description: "Rivals o1-mini on Math-500 and GPQA Diamond"
backend: "ollama"
ollama_name: "phi4:mini-reasoning"
llama_cpp_repo: "microsoft/Phi-4-mini-reasoning-gguf"
llama_cpp_file: "Phi-4-mini-reasoning-Q4_K_M.gguf"
params: 3.8e9
vram_gb: 3.0
ram_gb: 6.0
quantization: "Q4_K_M"
context_length: 128000
strengths:
- "reasoning"
- "math"
- "proofs"
- "chain-of-thought"
license: "MIT"
recommended_for:
- "math"
- "reasoning"
- "proofs"
release_date: "2025-05"
deepseek-r1-7b:
name: "DeepSeek-R1-Distill-Qwen-7B"
description: "Strong reasoning, distilled from R1"
backend: "ollama"
ollama_name: "deepseek-r1:7b"
llama_cpp_repo: "unsloth/DeepSeek-R1-Distill-Qwen-7B-GGUF"
llama_cpp_file: "DeepSeek-R1-Distill-Qwen-7B-Q4_K_M.gguf"
params: 7.0e9
vram_gb: 5.0
ram_gb: 8.0
quantization: "Q4_K_M"
context_length: 32768
strengths:
- "reasoning"
- "math"
- "code"
- "chain-of-thought"
license: "MIT"
recommended_for:
- "reasoning"
- "quality"
release_date: "2025-01"
# Multimodal capability
gemma-3n-e2b:
name: "Gemma-3n-E2B"
description: "Multimodal: text, image, audio, video"
backend: "ollama"
ollama_name: "gemma3:2b"
llama_cpp_repo: "google/gemma-3n-E2B-it-GGUF"
llama_cpp_file: "gemma-3n-E2B-it-Q4_K_M.gguf"
params: 5.0e9
effective_params: 2.0e9
vram_gb: 2.0
ram_gb: 4.0
quantization: "Q4_K_M"
context_length: 8192
multimodal: true
modalities:
- "text"
- "image"
- "audio"
- "video"
strengths:
- "multimodal"
- "vision"
- "audio"
license: "Apache-2.0"
recommended_for:
- "multimodal"
- "vision"
release_date: "2024-12"
phi-4-multimodal:
name: "Phi-4-multimodal"
description: "Microsoft multimodal with vision and speech"
backend: "ollama"
ollama_name: "phi4:multimodal"
llama_cpp_repo: "microsoft/Phi-4-multimodal-instruct-gguf"
llama_cpp_file: "Phi-4-multimodal-instruct-Q4_K_M.gguf"
params: 5.6e9
vram_gb: 4.0
ram_gb: 8.0
quantization: "Q4_K_M"
context_length: 128000
multimodal: true
modalities:
- "text"
- "image"
- "audio"
strengths:
- "multimodal"
- "vision"
- "speech"
- "reasoning"
license: "MIT"
recommended_for:
- "multimodal"
- "vision"
release_date: "2025-05"
# ============================================
# TIER 4: Cloud API Models (Requires API Key)
# Best for: Complex coding, agentic workflows
# ============================================
minimax-m2:
name: "MiniMax-M2"
description: "SOTA coding model - 230B MoE, 10B active params, optimized for SWE tasks"
backend: "openai_compatible"
api_base: "https://api.deepinfra.com/v1/openai"
api_model: "MiniMaxAI/MiniMax-M2"
params: 230e9
active_params: 10e9
context_length: 128000
interleaved_thinking: true
strengths:
- "code"
- "agentic"
- "multi-file-editing"
- "debugging"
- "swe-bench"
- "tool-calling"
license: "MiniMax"
recommended_for:
- "code"
- "agentic"
- "complex-coding"
env_key: "DEEPINFRA_API_KEY"
release_date: "2025-04"
notes: "Interleaved thinking model - preserve <think>...</think> tags in history"
minimax-m2-1:
name: "MiniMax-M2.1"
description: "Enhanced M2 with multi-language programming and office automation"
backend: "openai_compatible"
api_base: "https://api.deepinfra.com/v1/openai"
api_model: "MiniMaxAI/MiniMax-M2.1"
params: 230e9
active_params: 10e9
context_length: 128000
interleaved_thinking: true
strengths:
- "code"
- "agentic"
- "multi-language-code"
- "office-automation"
- "real-world-tasks"
license: "MiniMax"
recommended_for:
- "code"
- "real-world-dev"
env_key: "DEEPINFRA_API_KEY"
release_date: "2025-06"
# ============================================
# TIER 5: Vision/OCR Models (Document Processing)
# ============================================
deepseek-ocr2:
name: "DeepSeek-VL2"
description: "Advanced vision-language model for OCR and document parsing"
backend: "openai_compatible"
api_base: "https://api.deepinfra.com/v1/openai"
api_model: "deepseek-ai/deepseek-vl2"
params: 27e9
context_length: 32768
multimodal: true
modalities:
- "text"
- "image"
capabilities:
- "ocr"
- "document_parsing"
- "nutrition_labels"
- "lab_results"
- "prescription_reading"
strengths:
- "ocr"
- "vision"
- "document-understanding"
- "table-extraction"
- "handwriting"
license: "DeepSeek"
recommended_for:
- "document_parsing"
- "ocr"
- "health_documents"
env_key: "DEEPINFRA_API_KEY"
release_date: "2024-12"
notes: "Primary model for health document OCR - lab results, prescriptions, nutrition labels"
qwen-vl-max:
name: "Qwen-VL-Max"
description: "Alibaba's multimodal model for vision tasks"
backend: "openai_compatible"
api_base: "https://dashscope.aliyuncs.com/compatible-mode/v1"
api_model: "qwen-vl-max"
params: 72e9
context_length: 32768
multimodal: true
modalities:
- "text"
- "image"
capabilities:
- "ocr"
- "visual_qa"
- "document_analysis"
strengths:
- "vision"
- "ocr"
- "chinese"
- "document-understanding"
license: "Tongyi Qianwen"
recommended_for:
- "vision"
- "multilingual_ocr"
env_key: "DASHSCOPE_API_KEY"
release_date: "2024-08"
# Embedding models
embeddings:
all-minilm-l6-v2:
name: "all-MiniLM-L6-v2"
provider: "sentence-transformers"
dimensions: 384
max_sequence_length: 256
ram_mb: 100
recommended_for:
- "default"
- "speed"
bge-small-en:
name: "BAAI/bge-small-en-v1.5"
provider: "sentence-transformers"
dimensions: 384
max_sequence_length: 512
ram_mb: 150
recommended_for:
- "accuracy"
nomic-embed-text:
name: "nomic-embed-text"
provider: "ollama"
dimensions: 768
max_sequence_length: 8192
ram_mb: 300
recommended_for:
- "long-context"
gte-small:
name: "thenlper/gte-small"
provider: "sentence-transformers"
dimensions: 384
max_sequence_length: 512
ram_mb: 100
recommended_for:
- "retrieval"
# Hardware profiles for automatic model selection
hardware_profiles:
# MAX EVERYTHING - For users with massive hardware
max:
description: "MAX MODE: Unlimited resources, best quality, all features"
models:
- "minimax-m2"
- "phi-4-mini-reasoning"
- "deepseek-r1-7b"
- "phi-4-mini"
- "phi-4-multimodal"
- "qwen3-4b"
- "deepseek-r1-1.5b"
embedding: "nomic-embed-text"
swarm_enabled: true
swarm_strategy: "parallel_vote"
max_parallel: 8
enable_cloud: true
enable_multimodal: true
enable_speculation: true
context_window: 128000
notes: "Uses ALL available models including cloud APIs for maximum quality"
# Under 4GB total RAM
minimal:
description: "Minimal hardware (< 4GB RAM)"
models:
- "tinyllama-1.1b"
- "qwen3-0.6b"
embedding: "all-minilm-l6-v2"
swarm_enabled: false
# CPU only, 8GB+ RAM
cpu_only:
description: "CPU-only with 8GB+ RAM"
models:
- "bitnet-2b"
- "qwen3-0.6b"
- "tinyllama-1.1b"
- "smollm2-1.7b"
embedding: "all-minilm-l6-v2"
swarm_enabled: true
swarm_strategy: "fastest_first"
# 2-4 GB VRAM
low_vram:
description: "Low VRAM (2-4GB)"
models:
- "deepseek-r1-1.5b"
- "smollm2-1.7b"
- "qwen3-0.6b"
- "gemma-3n-e2b"
embedding: "all-minilm-l6-v2"
swarm_enabled: true
swarm_strategy: "mixture_of_experts"
# 4-8 GB VRAM
medium_vram:
description: "Medium VRAM (4-8GB)"
models:
- "phi-4-mini"
- "deepseek-r1-1.5b"
- "qwen3-4b"
- "smollm2-1.7b"
- "gemma-3n-e2b"
embedding: "bge-small-en"
swarm_enabled: true
swarm_strategy: "pso_collaborative"
# 8+ GB VRAM
high_vram:
description: "High VRAM (8GB+)"
models:
- "phi-4-mini-reasoning"
- "deepseek-r1-7b"
- "phi-4-mini"
- "qwen3-4b"
- "gemma-3n-e2b"
embedding: "nomic-embed-text"
swarm_enabled: true
swarm_strategy: "parallel_vote"
# Task-based model selection (Mixture of Experts routing)
task_routing:
code:
description: "Programming and debugging"
primary: "deepseek-r1-1.5b" # Local first - saves tokens
secondary: "phi-4-mini"
fallback: "qwen3-4b"
# Cloud escalation only for complex multi-file tasks
escalation_model: "minimax-m2"
escalation_threshold: 0.5 # Low confidence triggers cloud
swarm_models: # All local for swarm operations
- "deepseek-r1-1.5b"
- "phi-4-mini"
- "qwen3-4b"
notes: "Local models for swarm/memory, MiniMax M2 only for complex escalation"
reasoning:
description: "Logic and analysis"
primary: "phi-4-mini-reasoning"
secondary: "deepseek-r1-1.5b"
fallback: "qwen3-4b"
swarm_models:
- "phi-4-mini-reasoning"
- "deepseek-r1-1.5b"
- "phi-4-mini"
math:
description: "Mathematical computation"
primary: "phi-4-mini-reasoning"
secondary: "deepseek-r1-1.5b"
fallback: "qwen3-4b"
swarm_models:
- "phi-4-mini-reasoning"
- "deepseek-r1-1.5b"
creative:
description: "Writing and brainstorming"
primary: "qwen3-4b"
secondary: "smollm2-1.7b"
fallback: "qwen3-0.6b"
swarm_models:
- "qwen3-4b"
- "smollm2-1.7b"
research:
description: "Information gathering"
primary: "phi-4-mini"
secondary: "deepseek-r1-1.5b"
fallback: "smollm2-1.7b"
multilingual:
description: "Non-English languages"
primary: "qwen3-4b"
secondary: "qwen3-0.6b"
fallback: "gemma-3n-e2b"
multimodal:
description: "Vision, audio, video"
primary: "phi-4-multimodal"
secondary: "gemma-3n-e2b"
fallback: null
speed:
description: "Low-latency responses"
primary: "bitnet-2b"
secondary: "tinyllama-1.1b"
fallback: "qwen3-0.6b"
# Model swarm configurations
# ALL SWARMS USE LOCAL MODELS ONLY - saves tokens/context
swarm:
# Default swarm for general use
default:
description: "Balanced local swarm - saves tokens"
strategy: "mixture_of_experts"
use_local_only: true
models:
- id: "deepseek-r1-1.5b"
role: "reasoning"
weight: 1.0
- id: "phi-4-mini"
role: "quality"
weight: 0.8
- id: "qwen3-0.6b"
role: "speed"
weight: 0.5
- id: "smollm2-1.7b"
role: "generalist"
weight: 0.7
enable_verification: true
max_parallel: 3
# Cloud escalation for low-confidence results
cloud_escalation:
enabled: true
threshold: 0.4
model: "minimax-m2"
# Memory/context operations - ultra lightweight
memory:
description: "For memory retrieval, context building, summarization"
strategy: "fastest_first"
use_local_only: true
models:
- id: "qwen3-0.6b"
role: "summarize"
weight: 1.0
- id: "tinyllama-1.1b"
role: "speed"
weight: 0.9
enable_verification: false
max_parallel: 2
notes: "Tiny models for memory ops - preserves context window"
# Speed-optimized swarm
fast:
description: "Speed-optimized with fast escalation"
strategy: "fastest_first"
use_local_only: true
models:
- id: "bitnet-2b"
role: "speed"
weight: 1.0
- id: "tinyllama-1.1b"
role: "speed"
weight: 0.9
- id: "qwen3-0.6b"
role: "speed"
weight: 0.8
- id: "deepseek-r1-1.5b"
role: "reasoning"
weight: 0.6
enable_verification: false
max_parallel: 2
# Quality-optimized swarm
quality:
description: "Quality-first with verification"
strategy: "parallel_vote"
models:
- id: "phi-4-mini-reasoning"
role: "reasoning"
weight: 1.0
- id: "deepseek-r1-7b"
role: "reasoning"
weight: 0.9
- id: "phi-4-mini"
role: "quality"
weight: 0.8
- id: "qwen3-4b"
role: "generalist"
weight: 0.7
enable_verification: true
max_parallel: 4
# Reasoning-focused swarm
reasoning:
description: "Chain-of-thought reasoning specialists"
strategy: "speculative_ensemble"
models:
- id: "phi-4-mini-reasoning"
role: "reasoning"
weight: 1.0
- id: "deepseek-r1-1.5b"
role: "reasoning"
weight: 0.9
- id: "qwen3-0.6b"
role: "speed"
weight: 0.5
enable_verification: true
max_parallel: 2
# PSO collaborative swarm
collaborative:
description: "PSO-optimized collaborative inference"
strategy: "pso_collaborative"
pso_config:
inertia: 0.7
cognitive: 1.5
social: 1.5
models:
- id: "deepseek-r1-1.5b"
role: "reasoning"
- id: "phi-4-mini"
role: "quality"
- id: "smollm2-1.7b"
role: "generalist"
- id: "qwen3-0.6b"
role: "speed"
- id: "tinyllama-1.1b"
role: "speed"
enable_verification: true
max_parallel: 5
# MAX EVERYTHING - Ultimate swarm for massive hardware
max:
description: "MAXIMUM POWER - All models, parallel voting, cloud + local"
strategy: "parallel_vote"
use_local_only: false
enable_cloud: true
models:
- id: "minimax-m2"
role: "coding"
weight: 1.0
cloud: true
- id: "phi-4-mini-reasoning"
role: "reasoning"
weight: 1.0
- id: "deepseek-r1-7b"
role: "reasoning"
weight: 0.95
- id: "phi-4-mini"
role: "quality"
weight: 0.9
- id: "qwen3-4b"
role: "generalist"
weight: 0.85
- id: "deepseek-r1-1.5b"
role: "speed-reasoning"
weight: 0.8
- id: "phi-4-multimodal"
role: "multimodal"
weight: 0.8
enable_verification: true
verification_model: "phi-4-mini-reasoning"
max_parallel: 8
voting_threshold: 0.6
context_window: 128000
notes: "Run ALL models in parallel, vote on best response, verify with reasoning model"
# Speculative decoding pairs (draft + target)
speculative_decoding:
pairs:
- draft: "bitnet-2b"
target: "phi-4-mini"
speedup: 2.5
description: "BitNet drafts for Phi-4"
- draft: "tinyllama-1.1b"
target: "deepseek-r1-1.5b"
speedup: 1.8
description: "TinyLlama drafts for DeepSeek"
- draft: "qwen3-0.6b"
target: "qwen3-4b"
speedup: 2.0
description: "Qwen family speculative"
- draft: "smollm2-1.7b"
target: "phi-4-mini-reasoning"
speedup: 1.5
description: "SmolLM drafts for reasoning"
# Benchmark reference (for model selection decisions)
benchmarks:
# Scores are approximate, from various sources
phi-4-mini-reasoning:
math-500: 94.5
gpqa-diamond: 68.0
mmlu-pro: 72.0
deepseek-r1-1.5b:
math-500: 83.9
gpqa-diamond: 33.8
aime-2024: 28.3
phi-4-mini:
mmlu: 84.8
humaneval: 82.0
math: 80.4
qwen3-4b:
mmlu-pro: 74.0
gpqa-diamond: 67.0
livecode: 64.0
