AI / ML Models

AI / ML Models#

Reference of AI model families, model hosts, training datasets, and the evaluation benchmarks operators look at when picking a model.

For the broader NLP toolkit, see NLP. For data engineering around AI, see the “AI for NLP and data engineering” section there.

Family	Notes
GPT (OpenAI)	GPT-4 / 4o / o1 / o3 / o4 generations; multimodal; tool-use.
Claude (Anthropic)	3 Opus / Sonnet / Haiku / 3.5 / 3.7 / 4 / 4.5 / 4.6 / 4.7; strong on reasoning, code, and long context.
Gemini (Google)	Gemini 1.5 / 2.0 / 2.5 family; native multimodal.
Grok (xAI)	Grok 2 / 3 / 4.
Llama (Meta)	Llama 1 / 2 / 3 / 3.1 / 3.2 / 4 series; open weights.
Mistral / Mixtral	French OSS-friendly; Mistral 7B, Mixtral 8x7B / 8x22B, Mistral Large, Pixtral, Codestral.
Qwen (Alibaba)	open Chinese-origin; Qwen 2 / 2.5 / 3 / 3.5; strong coding + multilingual.
DeepSeek Yi (01.AI)	R1 (reasoning), V3 (chat); open weights.
Phi (Microsoft)	Phi-3 / 3.5 / 4; small + efficient open models.
Gemma (Google)	open Google models.
Command / Cohere	Command R+ retrieval-tuned.
Falcon (TII)	UAE-origin open.

Family	Notes
Stable Diffusion	1.x / 2.x / SDXL / 3 / 3.5; OSS image gen (Stability AI).
Flux	Black Forest Labs; open + paid tiers.
DALL-E 3	OpenAI; via ChatGPT.
Imagen 3 / Veo	Google; image / video.
Midjourney	v6 / v7 closed-platform.
Ideogram	text-strong gen.
Adobe Firefly	enterprise-friendly.
SAM 2 (Meta)	segment-anything; image + video.
Grounding DINO	open-vocabulary detection.
YOLO family	YOLOv8 / 9 / 10 / 11; real-time detection.
DETR / RT-DETR	transformer-based detection.
CLIP	contrastive language-image; the embedding backbone.
SigLIP / OpenCLIP	open variants.
DINOv2 / DINOv3	self-supervised vision transformers (Meta).

Family	Notes
Whisper (OpenAI)	multilingual ASR; large-v3 standard.
faster-whisper	CTranslate2-accelerated.
Distil-Whisper	smaller / faster variants.
SeamlessM4T (Meta)	ASR + speech-to-text + speech-to-speech.
NeMo (NVIDIA)	full speech + NLP + TTS.
Wav2Vec 2 / HuBERT	self-supervised acoustic.
Coqui TTS	OSS TTS; multi-voice.
ElevenLabs	commercial TTS / voice clone.
OpenVoice (MyShell)	OSS voice clone.
Piper	Linux-friendly local TTS.
Bark	OSS generative speech (Suno-AI).

Model	Notes
text-embedding-3 (O	penAI) small / large.
voyage-3 (Voyage AI	) commercial; strong retrieval scores.
cohere-embed-3 / 4	commercial.
gte (Alibaba)	OSS; strong scores per parameter.
bge (BAAI)	OSS; multilingual.
e5 (Microsoft)	OSS; strong on MTEB.
nomic-embed-text	OSS.
Stella, ColBERT, Co lPali	retrieval-specific.
multilingual-e5	MTEB-leading multilingual.

Model	Notes
GitHub Copilot	GPT-4-class backend; subscription.
Claude Code	CLI for code (Anthropic).
Cursor	IDE with GPT/Claude routing.
Codestral (Mistral)	open code model.
DeepSeek-Coder	open.
Qwen-Coder	open.
StarCoder 2	open (BigCode).
Code Llama	Meta open.
SWE-bench leaders	benchmark-leading code agents (Cognition Devin, Magic).

Platform	Notes
OpenAI API	GPT / Whisper / DALL-E / o-series.
Anthropic API	Claude.
Google AI Studio / Vertex AI	Gemini.
AWS Bedrock	multi-vendor (Claude, Llama, Mistral, Cohere, Titan).
Azure OpenAI	OpenAI on Azure.
Together AI	OSS hosting.
Replicate	OSS model hosting.
Fireworks AI	OSS hosting; OpenAI-compatible.
Anyscale	OSS hosting.
Groq	LPU-accelerated inference (Llama / Mixtral).
Cerebras	wafer-scale inference.
Cloudflare AI Worke	rs edge AI.
Hugging Face	largest model + dataset hub.
Ollama	local llm runtime.
LM Studio	local llm UI.
vLLM	GPU-optimized inference server.
Text Generation Inference (TGI)	Hugging Face inference server.
SGLang	modern inference framework.

Dataset	Notes
Common Crawl	petabyte web crawl since 2008; the standard large pretrain source.
C4	Google’s Colossal Clean Crawled Corpus.
RedPajama	open replication; Together AI.
The Pile (EleutherA	825GB diverse text.
RefinedWeb	Falcon-team’s clean Common Crawl.
SlimPajama	deduplicated RedPajama.
The Stack v2	BigCode source-code corpus.
StarCoder data	curated subset of The Stack.
Wikipedia / Wikidat	a classic high-quality.
Books3 / OpenBookCo	rpus / Project Gutenberg
ImageNet	classic vision benchmark + pretraining.
LAION-5B / 2B-en	image-text pairs; copyright-litigated.
JFT-300M / 3B	Google internal vision.
WebVid-10M	video.
Vimeo-CC, Panda-70M	, MSR-VTT video datasets.
LibriSpeech	classic ASR (1000 h English).
CommonVoice	Mozilla; multilingual ASR.
LibriLight, LibriHe avy	scaled ASR.

Benchmark	Notes
MMLU	multi-task language understanding (~14k qs over 57 subjects).
MMLU-Pro	harder MMLU.
GPQA Diamond	graduate-level science qs.
GSM8K	grade-school math word problems.
MATH	competition math.
HumanEval	Python code completion.
MBPP	Python code.
SWE-bench (Verified	) GitHub bug fixes.
LiveCodeBench	continuously refreshed code.
ARC-AGI	abstract reasoning (Chollet); hard for LLMs.
HellaSwag	commonsense.
BBH (BIG-Bench Hard	) reasoning subset.
LMSYS Chatbot Arena	human-preference Elo.
MT-Bench	multi-turn judge.
TruthfulQA	truthfulness.
HaluBench	hallucination detection.
CoreBench	judgment of expert tasks.
FRAMES, NIAH	long-context tests.
RULER	long-context with diverse tasks.
LongBench v2	long-context.
MTEB	massive text-embedding benchmark.
ImageNet, COCO, LVIS	vision.
LMMs-Eval, MMMU	multimodal LM eval.
WMT, FLORES	translation.

Accelerator	Notes
NVIDIA H100 / H200	Hopper; 80 GB / 141 GB HBM3.
NVIDIA B100 / B200	Blackwell; 192 GB HBM3e.
NVIDIA GB200 Grace Blackwell	superchip + ARM CPU.
NVIDIA A100	Ampere; 40 / 80 GB.
NVIDIA L40 / L40S	Ada Lovelace; data-center inference.
NVIDIA RTX 4090 / 5 090	consumer; 24 / 32 GB.
AMD MI300X / MI325X	192 / 256 GB.
Intel Gaudi 3	128 GB HBM2e.
Google TPU v5e / v5	p / Trillium / Ironwood (v6).
AWS Trainium 2 / In Azure Maia 100.	ferentia 2.
Groq LPU	cell-style accelerator.
Cerebras WSE-3 SambaNova SN40L.	wafer-scale.

Open weights vs. open source, Llama / Qwen / DeepSeek ship weights; license terms vary on commercial / extra-large use. “Open source” requires open data + code as well.
Quantisation, 4-bit / 8-bit GGUF / GPTQ / AWQ formats let 70B-class models run on consumer GPUs; quality cost is measurable but small.
Context length, 200K / 1M / 2M tokens are the 2026 norm for top models; needle-in-haystack works but multi-document reasoning still degrades past ~128K in many cases.
Tool use / agents, function calling + structured-output + persistent memory now standard; benchmarks (SWE-bench, WebArena, OSWorld) measure agent behavior.
Cost trends, per-token API prices have dropped 10-100× since 2023; capability per dollar improving fast.
Self-hosting break-even, somewhere around 1B-10B tokens / month makes self-hosting a 70B-class model cheaper than API.
Provenance + watermarking, C2PA, SynthID, generative- ai detection are emerging; not reliable yet.