CalvinBackup/OBLITERATUS
1
0
Fork 0
mirror of https://github.com/elder-plinius/OBLITERATUS.git synced 2026-08-18 00:47:23 +02:00
Code Issues Packages Projects Releases Wiki Activity
Files
dd75d5939fbbf6a64cef93fcdeddf64eb488946d
OBLITERATUS/obliteratus/presets.py
T
pliny 0f6114fe87 Add files via upload
2026-03-04 12:38:18 -08:00

1183 lines
50 KiB
Python
Raw Blame History

"""Model presets organized by provider, ordered by model size.
Each model is tagged with a compute tier for hardware auto-detection:
- tiny: Runs on any machine, even CPU-only laptops (< 1GB VRAM/RAM)
- small: Needs ~4GB VRAM or 8GB RAM (a basic GPU or CPU with patience)
- medium: Needs ~8-16GB VRAM (consumer GPU like RTX 3060/4060)
- large: Needs 24GB+ VRAM (RTX 3090/4090 or A100)
- frontier: Multi-GPU or cloud. Top LM Arena open-weight models (MoE/dense 70B+)
"""
from __future__ import annotations
from dataclasses import dataclass
@dataclass
class ModelPreset:
name: str
hf_id: str
description: str
tier: str # tiny, small, medium, large
params: str # human-readable param count
recommended_dtype: str
recommended_quantization: str | None = None # "4bit", "8bit", or None
gated: bool = False # True if the HF repo requires license acceptance + token
# Curated list of popular open-source models across compute tiers
MODEL_PRESETS: dict[str, ModelPreset] = {}
_PRESETS_LIST = [
# ╔══════════════════════════════════════════════════════════════════╗
# ║ 01.AI (Yi) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="Yi 1.5 6B Chat",
hf_id="01-ai/Yi-1.5-6B-Chat",
description="01.AI's Yi 1.5 6B. Bilingual EN/ZH, strong math and coding. Apache 2.0.",
tier="medium",
params="6B",
recommended_dtype="float16",
),
ModelPreset(
name="Yi 1.5 9B Chat",
hf_id="01-ai/Yi-1.5-9B-Chat",
description="01.AI's Yi 1.5 9B. Top-tier bilingual model, beats GPT-3.5 on benchmarks.",
tier="medium",
params="9B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Yi 1.5 34B Chat",
hf_id="01-ai/Yi-1.5-34B-Chat",
description="01.AI's flagship 34B model. Strong multilingual reasoning. Apache 2.0.",
tier="large",
params="34B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Alibaba Cloud (Qwen) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="Qwen2.5-0.5B",
hf_id="Qwen/Qwen2.5-0.5B",
description="Tiny Qwen model, very fast ablation studies.",
tier="tiny",
params="0.5B",
recommended_dtype="float16",
),
ModelPreset(
name="Qwen2.5-0.5B Instruct",
hf_id="Qwen/Qwen2.5-0.5B-Instruct",
description="Tiny Qwen instruct model, fast ablation studies with chat template.",
tier="tiny",
params="0.5B",
recommended_dtype="float16",
),
ModelPreset(
name="Qwen3-0.6B",
hf_id="Qwen/Qwen3-0.6B",
description="Qwen3 0.6B — smallest Qwen3 with think/non-think modes.",
tier="tiny",
params="0.6B",
recommended_dtype="float16",
),
ModelPreset(
name="Qwen2.5-1.5B",
hf_id="Qwen/Qwen2.5-1.5B",
description="Qwen 1.5B — strong multilingual model.",
tier="small",
params="1.5B",
recommended_dtype="float16",
),
ModelPreset(
name="Qwen2.5-1.5B Instruct",
hf_id="Qwen/Qwen2.5-1.5B-Instruct",
description="Qwen 1.5B instruct — strong multilingual chat model.",
tier="small",
params="1.5B",
recommended_dtype="float16",
),
ModelPreset(
name="Qwen3-1.7B",
hf_id="Qwen/Qwen3-1.7B",
description="Qwen3 1.7B — compact Qwen3 with think/non-think modes.",
tier="small",
params="1.7B",
recommended_dtype="float16",
),
ModelPreset(
name="Qwen2.5-3B Instruct",
hf_id="Qwen/Qwen2.5-3B-Instruct",
description="Qwen 3B instruct — excellent small chat model.",
tier="small",
params="3B",
recommended_dtype="float16",
),
ModelPreset(
name="Qwen3-4B",
hf_id="Qwen/Qwen3-4B",
description="Qwen3 4B — strong reasoning with think/non-think modes. Apache 2.0.",
tier="small",
params="4B",
recommended_dtype="float16",
),
ModelPreset(
name="Qwen2.5-7B",
hf_id="Qwen/Qwen2.5-7B",
description="Strong 7B Qwen model. Use 4-bit quantization on 8GB GPUs.",
tier="medium",
params="7B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Qwen2.5-7B Instruct",
hf_id="Qwen/Qwen2.5-7B-Instruct",
description="Qwen 7B instruct variant with chat template.",
tier="medium",
params="7B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Qwen2.5 Coder 7B Instruct",
hf_id="Qwen/Qwen2.5-Coder-7B-Instruct",
description="Qwen 7B fine-tuned for code generation and understanding.",
tier="medium",
params="7B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Qwen3-8B",
hf_id="Qwen/Qwen3-8B",
description="Qwen3 8B — strong reasoning, think/non-think modes. Apache 2.0.",
tier="medium",
params="8B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Qwen2.5-14B",
hf_id="Qwen/Qwen2.5-14B",
description="Qwen 14B — needs quantization for consumer GPUs.",
tier="large",
params="14B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Qwen2.5-72B",
hf_id="Qwen/Qwen2.5-72B",
description="Qwen 72B dense — frontier-class model. Multi-GPU. Apache 2.0.",
tier="frontier",
params="72B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="Qwen3-32B",
hf_id="Qwen/Qwen3-32B",
description="Qwen 32B — frontier-class open model. Multi-GPU or heavy quant.",
tier="large",
params="32B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="Qwen3 30B-A3B",
hf_id="Qwen/Qwen3-30B-A3B",
description="Qwen3 MoE — 30B total, 3B active. Runs on consumer GPU. Think/non-think modes.",
tier="large",
params="30B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="Qwen3 235B-A22B",
hf_id="Qwen/Qwen3-235B-A22B",
description="Qwen3 flagship. 235B MoE (22B active), 128 experts. Think/non-think. Apache 2.0.",
tier="frontier",
params="235B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="Qwen3.5-0.8B",
hf_id="Qwen/Qwen3.5-0.8B",
description="Qwen3.5 0.8B — Gated DeltaNet hybrid attention, 262K ctx. Non-thinking mode. Apache 2.0.",
tier="tiny",
params="0.8B",
recommended_dtype="float16",
),
ModelPreset(
name="Qwen3.5-2B",
hf_id="Qwen/Qwen3.5-2B",
description="Qwen3.5 2B — Gated DeltaNet hybrid attention, 262K ctx, think/non-think modes. Apache 2.0.",
tier="small",
params="2B",
recommended_dtype="float16",
),
ModelPreset(
name="Qwen3.5-4B",
hf_id="Qwen/Qwen3.5-4B",
description="Qwen3.5 4B — natively multimodal, 262K ctx, think/non-think modes. Apache 2.0.",
tier="small",
params="4B",
recommended_dtype="float16",
),
ModelPreset(
name="Qwen3.5-9B",
hf_id="Qwen/Qwen3.5-9B",
description="Qwen3.5 9B — natively multimodal, 262K ctx (1M extended), think/non-think modes. Apache 2.0.",
tier="medium",
params="9B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Qwen3.5 397B-A17B",
hf_id="Qwen/Qwen3.5-397B-A17B",
description="Qwen3.5 flagship. 397B MoE (17B active). Gated DeltaNet hybrid attention. 262K ctx. Apache 2.0.",
tier="frontier",
params="397B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Allen Institute for AI (AI2) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="OLMo 2 1B Instruct",
hf_id="allenai/OLMo-2-0425-1B-Instruct",
description="AI2's compact fully open 1B instruct model. Apache 2.0.",
tier="tiny",
params="1B",
recommended_dtype="float16",
),
ModelPreset(
name="OLMo 3 7B Instruct",
hf_id="allenai/Olmo-3-7B-Instruct",
description="AI2's fully open 7B instruct model. Apache 2.0.",
tier="medium",
params="7B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="OLMo 2 32B Instruct",
hf_id="allenai/OLMo-2-0325-32B-Instruct",
description="AI2's fully open 32B model (data+code+weights). Apache 2.0.",
tier="large",
params="32B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="OLMo 3.1 32B Think",
hf_id="allenai/Olmo-3.1-32B-Think",
description="AI2's fully open model (data+code+weights). Chain-of-thought reasoning. Apache 2.0.",
tier="large",
params="32B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Apple ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="OpenELM 270M",
hf_id="apple/OpenELM-270M",
description="Apple's efficient 270M LM. Layer-wise scaling, fully open training recipe.",
tier="tiny",
params="270M",
recommended_dtype="float32",
),
ModelPreset(
name="OpenELM 1.1B",
hf_id="apple/OpenELM-1_1B",
description="Apple's 1.1B model. Layer-wise scaling for efficient inference.",
tier="tiny",
params="1.1B",
recommended_dtype="float16",
),
ModelPreset(
name="OpenELM 3B",
hf_id="apple/OpenELM-3B",
description="Apple's largest OpenELM. Competitive at 3B scale with open training data.",
tier="small",
params="3B",
recommended_dtype="float16",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Cohere For AI ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="Aya Expanse 8B",
hf_id="CohereForAI/aya-expanse-8b",
description="Cohere's massively multilingual model. 23 languages. Apache 2.0.",
tier="medium",
params="8B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Aya Expanse 32B",
hf_id="CohereForAI/aya-expanse-32b",
description="Cohere's flagship multilingual model. 23 languages. Apache 2.0.",
tier="large",
params="32B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="Command R",
hf_id="CohereForAI/c4ai-command-r-v01",
description="Cohere's 35B RAG-optimized model. 10 languages, 128K context. CC-BY-NC.",
tier="large",
params="35B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="Command R+",
hf_id="CohereForAI/c4ai-command-r-plus",
description="Cohere's 104B flagship. Tool use, RAG, multilingual. CC-BY-NC.",
tier="frontier",
params="104B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Databricks ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="DBRX Instruct",
hf_id="databricks/dbrx-instruct",
description="132B MoE (36B active, 16 experts). Strong code + reasoning. Databricks Open.",
tier="frontier",
params="132B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ DeepSeek ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="DeepSeek-R1 Distill Qwen 1.5B",
hf_id="deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B",
description="DeepSeek-R1 reasoning distilled into Qwen 1.5B. Chain-of-thought in a tiny model. MIT.",
tier="small",
params="1.5B",
recommended_dtype="float16",
),
ModelPreset(
name="DeepSeek-R1 Distill Qwen 7B",
hf_id="deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
description="DeepSeek-R1 reasoning distilled into Qwen 7B. Strong chain-of-thought. MIT.",
tier="medium",
params="7B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="DeepSeek-R1 Distill Llama 8B",
hf_id="deepseek-ai/DeepSeek-R1-Distill-Llama-8B",
description="DeepSeek-R1 reasoning distilled into Llama 8B. Best reasoning at 8B scale. MIT.",
tier="medium",
params="8B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="DeepSeek-R1 Distill Qwen 14B",
hf_id="deepseek-ai/DeepSeek-R1-Distill-Qwen-14B",
description="DeepSeek-R1 reasoning distilled into Qwen 14B. Top mid-range reasoner. MIT.",
tier="large",
params="14B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="DeepSeek-R1 Distill Qwen 32B",
hf_id="deepseek-ai/DeepSeek-R1-Distill-Qwen-32B",
description="DeepSeek-R1 reasoning distilled into Qwen 32B. Strong chain-of-thought. MIT license.",
tier="large",
params="32B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="DeepSeek-R1 Distill Llama 70B",
hf_id="deepseek-ai/DeepSeek-R1-Distill-Llama-70B",
description="DeepSeek-R1 reasoning distilled into Llama 70B. Near-frontier reasoning. MIT license.",
tier="large",
params="70B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="DeepSeek-R1 0528 Qwen3-8B",
hf_id="deepseek-ai/DeepSeek-R1-0528-Qwen3-8B",
description="DeepSeek-R1 reasoning distilled into Qwen3 8B. Latest R1 distillation. MIT.",
tier="medium",
params="8B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="DeepSeek-V3",
hf_id="deepseek-ai/DeepSeek-V3",
description="685B MoE (37B active). MLA + DeepSeekMoE. MIT license.",
tier="frontier",
params="685B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="DeepSeek-R1",
hf_id="deepseek-ai/DeepSeek-R1",
description="671B MoE reasoning model. RL-trained chain-of-thought. MIT license.",
tier="frontier",
params="671B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ EleutherAI (Pythia) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="Pythia 160M",
hf_id="EleutherAI/pythia-160m",
description="EleutherAI's 160M model. Fully open training data (The Pile). Apache 2.0.",
tier="tiny",
params="160M",
recommended_dtype="float32",
),
ModelPreset(
name="Pythia 410M",
hf_id="EleutherAI/pythia-410m",
description="EleutherAI's 410M param model. Fully open training data (The Pile).",
tier="tiny",
params="410M",
recommended_dtype="float32",
),
ModelPreset(
name="Pythia 1.4B",
hf_id="EleutherAI/pythia-1.4b",
description="EleutherAI's 1.4B model. Fully reproducible, open data. Apache 2.0.",
tier="small",
params="1.4B",
recommended_dtype="float16",
),
ModelPreset(
name="Pythia 2.8B",
hf_id="EleutherAI/pythia-2.8b",
description="EleutherAI's 2.8B model. Fully reproducible, open data (The Pile). Apache 2.0.",
tier="small",
params="2.8B",
recommended_dtype="float16",
),
ModelPreset(
name="Pythia 6.9B",
hf_id="EleutherAI/pythia-6.9b",
description="EleutherAI's 6.9B model. Large-scale open research model. Apache 2.0.",
tier="medium",
params="6.9B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Pythia 12B",
hf_id="EleutherAI/pythia-12b",
description="EleutherAI's largest Pythia. Full suite of training checkpoints. Apache 2.0.",
tier="large",
params="12B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Google (Gemma) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="Gemma-2 2B",
hf_id="google/gemma-2-2b",
description="Google's compact Gemma model. Modern architecture. Requires HF approval.",
tier="small",
params="2B",
recommended_dtype="float16",
gated=True,
),
ModelPreset(
name="Gemma-2 9B",
hf_id="google/gemma-2-9b",
description="Google's 9B Gemma. Excellent for ablation at scale. Requires HF approval.",
tier="medium",
params="9B",
recommended_dtype="float16",
recommended_quantization="4bit",
gated=True,
),
ModelPreset(
name="Gemma 3 12B",
hf_id="google/gemma-3-12b-it",
description="Google's Gemma 3 12B. Multimodal, 128K context, 140+ languages. Requires HF approval.",
tier="medium",
params="12B",
recommended_dtype="float16",
recommended_quantization="4bit",
gated=True,
),
ModelPreset(
name="Gemma 3 27B",
hf_id="google/gemma-3-27b-it",
description="Google's Gemma 3 27B. Beats Gemini 1.5 Pro. Multimodal, 128K context, 140+ languages. Requires HF approval.",
tier="large",
params="27B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
gated=True,
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ HuggingFace ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="DistilGPT-2",
hf_id="distilbert/distilgpt2",
description="Distilled GPT-2 — only 82M params. Fastest option.",
tier="tiny",
params="82M",
recommended_dtype="float32",
),
ModelPreset(
name="SmolLM2-135M",
hf_id="HuggingFaceTB/SmolLM2-135M",
description="Extremely small modern LM. Great for quick iteration.",
tier="tiny",
params="135M",
recommended_dtype="float32",
),
ModelPreset(
name="SmolLM2-360M",
hf_id="HuggingFaceTB/SmolLM2-360M",
description="HuggingFace's 360M model. Efficient on-device inference.",
tier="tiny",
params="360M",
recommended_dtype="float32",
),
ModelPreset(
name="SmolLM2-1.7B",
hf_id="HuggingFaceTB/SmolLM2-1.7B-Instruct",
description="HuggingFace's 1.7B instruct model. Strong for its size.",
tier="small",
params="1.7B",
recommended_dtype="float16",
),
ModelPreset(
name="SmolLM3-3B",
hf_id="HuggingFaceTB/SmolLM3-3B",
description="HuggingFace's SmolLM3 3B. Latest efficient small LM.",
tier="small",
params="3B",
recommended_dtype="float16",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ IBM (Granite) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="Granite 3.1 2B Instruct",
hf_id="ibm-granite/granite-3.1-2b-instruct",
description="IBM's Granite 3.1 2B. Enterprise-grade, multilingual, Apache 2.0.",
tier="small",
params="2B",
recommended_dtype="float16",
),
ModelPreset(
name="Granite 3.1 8B Instruct",
hf_id="ibm-granite/granite-3.1-8b-instruct",
description="IBM's Granite 3.1 8B. Enterprise-focused, strong tool use. Apache 2.0.",
tier="medium",
params="8B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Meta (LLaMA) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="TinyLlama 1.1B",
hf_id="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
description="Compact LLaMA architecture, great for testing LLaMA-family ablation.",
tier="tiny",
params="1.1B",
recommended_dtype="float16",
),
ModelPreset(
name="LLaMA-3.1 8B",
hf_id="meta-llama/Llama-3.1-8B",
description="Meta's LLaMA 3.1. Requires HF approval.",
tier="large",
params="8B",
recommended_dtype="float16",
recommended_quantization="4bit",
gated=True,
),
ModelPreset(
name="LLaMA-3.1 70B",
hf_id="meta-llama/Llama-3.1-70B",
description="Meta's LLaMA 3.1 70B. Top dense open model. Requires HF approval.",
tier="frontier",
params="70B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
gated=True,
),
ModelPreset(
name="Llama 4 Scout",
hf_id="meta-llama/Llama-4-Scout-17B-16E-Instruct",
description="Meta's 109B MoE (17B active). 10M token context window. Multimodal. Requires HF approval.",
tier="frontier",
params="109B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
gated=True,
),
ModelPreset(
name="Llama 4 Maverick",
hf_id="meta-llama/Llama-4-Maverick-17B-128E-Instruct",
description="Meta's ~400B MoE (17B active, 128 experts). 1M ctx. Multimodal. 200 languages. Requires HF approval.",
tier="frontier",
params="400B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
gated=True,
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Microsoft (Phi) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="Phi-2",
hf_id="microsoft/phi-2",
description="Microsoft's 2.7B param model. Punches above its weight.",
tier="small",
params="2.7B",
recommended_dtype="float16",
),
ModelPreset(
name="Phi-3.5 Mini",
hf_id="microsoft/Phi-3.5-mini-instruct",
description="Microsoft's 3.8B param Phi-3.5. Great performance/size ratio.",
tier="medium",
params="3.8B",
recommended_dtype="float16",
),
ModelPreset(
name="Phi-4 Mini Instruct",
hf_id="microsoft/Phi-4-mini-instruct",
description="Microsoft's 3.8B Phi-4 Mini. Strong reasoning for its size. MIT license.",
tier="small",
params="3.8B",
recommended_dtype="float16",
),
ModelPreset(
name="Phi-4",
hf_id="microsoft/phi-4",
description="Microsoft's 14B Phi-4. State-of-the-art at its size. Requires HF approval.",
tier="large",
params="14B",
recommended_dtype="float16",
recommended_quantization="4bit",
gated=True,
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ MiniMax ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="MiniMax M2.1",
hf_id="MiniMaxAI/MiniMax-M2.1",
description="230B MoE (10B active). #1 open-source on Artificial Analysis composite. Modified-MIT.",
tier="frontier",
params="230B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="MiniMax M2.5",
hf_id="MiniMaxAI/MiniMax-M2.5",
description="230B MoE (10B active). Lightning Attention hybrid (7:1). CISPO RL-tuned. Modified-MIT.",
tier="frontier",
params="230B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Mistral AI ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="Mistral 7B v0.3",
hf_id="mistralai/Mistral-7B-v0.3",
description="Mistral's 7B model. Widely studied architecture. Requires HF approval.",
tier="medium",
params="7B",
recommended_dtype="float16",
recommended_quantization="4bit",
gated=True,
),
ModelPreset(
name="Mistral Nemo 12B",
hf_id="mistralai/Mistral-Nemo-Instruct-2407",
description="Mistral-NVIDIA collab. 12B dense, 128K context. Requires HF approval.",
tier="medium",
params="12B",
recommended_dtype="float16",
recommended_quantization="4bit",
gated=True,
),
ModelPreset(
name="Mistral Small 24B",
hf_id="mistralai/Mistral-Small-24B-Instruct-2501",
description="Mistral's 24B model. Strong reasoning, needs quantization. Requires HF approval.",
tier="large",
params="24B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
gated=True,
),
ModelPreset(
name="Mistral Small 3.1 24B",
hf_id="mistralai/Mistral-Small-3.1-24B-Instruct-2503",
description="Mistral Small 3.1 — vision + 128K context in a compact dense model. Requires HF approval.",
tier="large",
params="24B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
gated=True,
),
ModelPreset(
name="Mistral Large 3",
hf_id="mistralai/Mistral-Large-3-675B-Instruct-2512",
description="675B MoE (41B active). Vision + 256K ctx. Best agentic capabilities. Requires HF approval.",
tier="frontier",
params="675B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
gated=True,
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ OpenBMB ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="MiniCPM3-4B",
hf_id="openbmb/MiniCPM3-4B",
description="OpenBMB's MiniCPM3 4B. Efficient on-device LM with strong reasoning.",
tier="small",
params="4B",
recommended_dtype="float16",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Moonshot AI (Kimi) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="Kimi-K2 Instruct",
hf_id="moonshotai/Kimi-K2-Instruct",
description="Moonshot's Kimi-K2 MoE model. 1T total params, ~32B active. Use trust_remote_code.",
tier="large",
params="1T MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="Kimi K2.5",
hf_id="moonshotai/Kimi-K2.5",
description="Moonshot's 1T MoE (32B active). Top coding + reasoning. 256K multimodal context.",
tier="frontier",
params="1T MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ NVIDIA ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="Nemotron Mini 4B",
hf_id="nvidia/Nemotron-Mini-4B-Instruct",
description="NVIDIA's compact 4B model. Optimized for on-device deployment.",
tier="small",
params="4B",
recommended_dtype="float16",
),
ModelPreset(
name="Mistral NeMo Minitron 8B",
hf_id="nvidia/Mistral-NeMo-Minitron-8B-Instruct",
description="NVIDIA pruned + distilled Mistral NeMo. 8B with 12B-level quality.",
tier="medium",
params="8B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Llama-3.1 Nemotron 70B",
hf_id="nvidia/Llama-3.1-Nemotron-70B-Instruct-HF",
description="NVIDIA's RLHF-tuned Llama 70B. Top reward model quality. Custom NVIDIA license.",
tier="frontier",
params="70B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ OpenAI ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="GPT-OSS 20B",
hf_id="openai/gpt-oss-20b",
description="OpenAI's first open-weight MoE (20B total, 3.6B active). MIT license.",
tier="large",
params="20B MoE",
recommended_dtype="float16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ OpenAI Community (GPT-2) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="GPT-2 Small",
hf_id="openai-community/gpt2",
description="Classic 124M param model. Perfect for learning and quick experiments.",
tier="tiny",
params="124M",
recommended_dtype="float32",
),
ModelPreset(
name="GPT-2 Medium",
hf_id="openai-community/gpt2-medium",
description="355M param GPT-2 variant. Good balance of size and capability.",
tier="tiny",
params="355M",
recommended_dtype="float32",
),
ModelPreset(
name="GPT-2 Large",
hf_id="openai-community/gpt2-large",
description="774M param GPT-2. Good for detailed layer ablation studies.",
tier="small",
params="774M",
recommended_dtype="float16",
),
ModelPreset(
name="GPT-2 XL",
hf_id="openai-community/gpt2-xl",
description="1.5B param GPT-2. Largest GPT-2 variant.",
tier="small",
params="1.5B",
recommended_dtype="float16",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Shanghai AI Lab (InternLM) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="InternLM3-8B Instruct",
hf_id="internlm/internlm3-8b-instruct",
description="Shanghai AI Lab's InternLM3 8B instruct. Strong reasoning. Apache 2.0.",
tier="medium",
params="8B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="InternLM2.5 7B Chat",
hf_id="internlm/internlm2_5-7b-chat",
description="Shanghai AI Lab's 7B model. Strong reasoning, 1M context. Apache 2.0.",
tier="medium",
params="7B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="InternLM2.5 20B Chat",
hf_id="internlm/internlm2_5-20b-chat",
description="Shanghai AI Lab's 20B model. Top-tier multilingual reasoning. Apache 2.0.",
tier="large",
params="20B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Stability AI (StableLM) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="StableLM-2 1.6B",
hf_id="stabilityai/stablelm-2-1_6b",
description="Stability AI's compact LM.",
tier="small",
params="1.6B",
recommended_dtype="float16",
),
ModelPreset(
name="StableLM-2 Zephyr 1.6B",
hf_id="stabilityai/stablelm-2-zephyr-1_6b",
description="Stability AI's chat-tuned StableLM. Direct Preference Optimization.",
tier="small",
params="1.6B",
recommended_dtype="float16",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ StepFun ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="Step 3.5 Flash",
hf_id="stepfun-ai/Step-3.5-Flash",
description="197B MoE (11B active). 100-350 tok/s. Strong on benchmarks. Apache 2.0.",
tier="frontier",
params="197B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Technology Innovation Institute (Falcon) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="Falcon3-3B Instruct",
hf_id="tiiuae/Falcon3-3B-Instruct",
description="TII's Falcon3 3B instruct. Modern architecture, Apache 2.0.",
tier="small",
params="3B",
recommended_dtype="float16",
),
ModelPreset(
name="Falcon 7B",
hf_id="tiiuae/falcon-7b",
description="TII's 7B model. Trained on 1.5T tokens of RefinedWeb. Apache 2.0.",
tier="medium",
params="7B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Falcon3-7B Instruct",
hf_id="tiiuae/Falcon3-7B-Instruct",
description="TII's Falcon3 7B instruct. Modern architecture, Apache 2.0.",
tier="medium",
params="7B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Falcon 11B",
hf_id="tiiuae/falcon-11B",
description="TII's 11B model. Improved architecture over Falcon 7B/40B. Apache 2.0.",
tier="medium",
params="11B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Falcon 40B",
hf_id="tiiuae/falcon-40b",
description="TII's 40B model. Strong multilingual capabilities. Apache 2.0.",
tier="large",
params="40B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Zhipu AI / THUDM (GLM) ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="GLM-4 9B",
hf_id="THUDM/glm-4-9b",
description="Tsinghua's GLM-4 9B. Bilingual (EN/ZH), strong reasoning.",
tier="medium",
params="9B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="GLM-4 9B Chat",
hf_id="THUDM/glm-4-9b-chat",
description="GLM-4 9B chat variant. Bilingual EN/ZH with tool calling.",
tier="large",
params="9B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="GLM-4 9B Chat HF",
hf_id="THUDM/glm-4-9b-chat-hf",
description="GLM-4 9B chat variant (HuggingFace-compatible format). No trust_remote_code needed.",
tier="medium",
params="9B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="GLM-4 32B Chat",
hf_id="zai-org/GLM-4-32B-0414",
description="GLM-4 32B. Strong bilingual EN/ZH with tool-calling. MIT license.",
tier="large",
params="32B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="GLM-4.7 Flash",
hf_id="zai-org/GLM-4.7-Flash",
description="GLM-4.7 Flash MoE — 30B total, 3B active. Runs on consumer GPU. MIT.",
tier="large",
params="30B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="GLM-4.7",
hf_id="zai-org/GLM-4.7",
description="#1 open-weight on LM Arena. 355B MoE (32B active). MIT. Thinking modes, 200K ctx.",
tier="frontier",
params="355B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
ModelPreset(
name="GLM-5",
hf_id="zai-org/GLM-5",
description="744B MoE (40B active). DeepSeek Sparse Attention + MLA. 200K ctx. MIT license.",
tier="frontier",
params="744B MoE",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Community Fine-tunes: Uncensored / Abliterated ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="Qwen2.5-7B Abliterated",
hf_id="huihui-ai/Qwen2.5-7B-Instruct-abliterated",
description="Qwen 7B with refusal direction removed. Compare vs. base for alignment research.",
tier="medium",
params="7B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Dolphin 2.9 Llama-3.1 8B",
hf_id="cognitivecomputations/dolphin-2.9.4-llama3.1-8b",
description="Uncensored Dolphin fine-tune. No alignment filtering. Popular for research.",
tier="medium",
params="8B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Hermes 3 Llama-3.1 8B",
hf_id="NousResearch/Hermes-3-Llama-3.1-8B",
description="Nous Hermes 3 — uncensored research model with strong reasoning.",
tier="medium",
params="8B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Llama-3.1 8B Abliterated",
hf_id="mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated",
description="LLaMA 3.1 with refusal direction abliterated. A/B test vs. base for jailbreak research.",
tier="large",
params="8B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Llama-3.1 8B Lexi Uncensored",
hf_id="Orenguteng/Llama-3.1-8B-Lexi-Uncensored-V2",
description="Fully uncensored LLaMA 3.1 fine-tune. No refusal training.",
tier="large",
params="8B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="Dolphin 2.9 Mistral 24B",
hf_id="cognitivecomputations/dolphin-2.9.4-mistral-24b",
description="Uncensored Dolphin on Mistral 24B base. Powerful unfiltered reasoning.",
tier="large",
params="24B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
# ╔══════════════════════════════════════════════════════════════════╗
# ║ Community Fine-tunes: Cybersecurity ║
# ╚══════════════════════════════════════════════════════════════════╝
ModelPreset(
name="WhiteRabbitNeo 7B",
hf_id="WhiteRabbitNeo/WhiteRabbitNeo-2.5-Qwen-2.5-Coder-7B",
description="Cybersecurity-focused model. Pentesting, exploit analysis, CTF.",
tier="medium",
params="7B",
recommended_dtype="float16",
recommended_quantization="4bit",
),
ModelPreset(
name="WhiteRabbitNeo 33B",
hf_id="WhiteRabbitNeo/WhiteRabbitNeo-33B-DeepSeekCoder",
description="Large cybersecurity model. Vuln analysis, exploit dev, red-teaming.",
tier="large",
params="33B",
recommended_dtype="bfloat16",
recommended_quantization="4bit",
),
]
for p in _PRESETS_LIST:
MODEL_PRESETS[p.hf_id] = p
def get_presets_by_tier(tier: str) -> list[ModelPreset]:
"""Return all presets for a compute tier."""
return [p for p in MODEL_PRESETS.values() if p.tier == tier]
def list_all_presets() -> list[ModelPreset]:
"""Return all presets sorted by tier then name."""
tier_order = {"tiny": 0, "small": 1, "medium": 2, "large": 3, "frontier": 4}
return sorted(MODEL_PRESETS.values(), key=lambda p: (tier_order.get(p.tier, 99), p.name))
def is_gated(hf_id: str) -> bool:
"""Return True if the given HF model ID is a known gated repo."""
preset = MODEL_PRESETS.get(hf_id)
if preset is not None:
return preset.gated
# Heuristic for custom model IDs: check common gated org prefixes
_GATED_ORGS = {"meta-llama", "google", "mistralai"}
org = hf_id.split("/")[0] if "/" in hf_id else ""
return org in _GATED_ORGS
Reference in New Issue View Git Blame Copy Permalink