Back Matter · Reference

The Reference

Every figure in this guide was checked against the source listed for it on 20 September 2026. Where I could find only a vendor's own claim, the text says "reported". Where I could not find a source, the figure is not here.

The Index

Each entry names the tier and the section that places the term.

A2A: Apex, What gets bolted on
Abliterated: Apex, What did someone do to the base
Active parameters: Vector, What does "235B-A22B" mean
Adapter: Vector, What is a checkpoint
Agent, harness, subagent, skill, hook, sandbox: Origin, Where do you meet one; Apex, What gets bolted on
Agentic, workflow: Apex, What gets bolted on
Alignment, guardrails: Apex, How was it made
API: Origin, Where do you meet one; Vector, What is inference
API-only: Vector, Are open weights open source
Arena, benchmark, contamination: Apex, How is it evaluated
Artificial intelligence, AI: Intro, What is "artificial intelligence"; Origin, What does "AI" actually mean on the box
ASR, STT, WER: Nexus, Speech and audio
Attention (GQA, MLA, sliding window, linear, DeltaNet): Apex, What do the architecture words on a release post mean
AWQ, GPTQ, FP8, NVFP4, MXFP4: Vector, Why does one model have thirty downloads; Nexus, Language models: which variant

Base, Instruct, Chat, Thinking: Nexus, Language models: which variant
Batch (as a speed): Vector, What is inference; (as an API): Apex, What are the words on the invoice
Benchmark table, pass@1, n-shot: Apex, How do you read the table
BF16, FP16, INT8, INT4: Vector, Why does one model have thirty downloads
Bit, byte, gigabyte: Origin, Why is it a download
Borrowed words (learn, know, understand, think, reason, remember): Origin, How did it learn
Bytes per parameter: Vector, Why does one model have thirty downloads

C2PA, watermark, SynthID: Apex, How is it evaluated
Calibration, confidence: Apex, What gets bolted on; How was it made
Cap, plan: Origin, What is it actually reading
CFG, guidance, sampler, seed, negative prompt: Nexus, Image generation
Chain of thought, few-shot: Apex, What gets bolted on
Chat product: Origin, Where do you meet one
Chat template, special tokens: Vector, What is in the folder
Chatbot: Origin, Where do you meet one
Checkpoint: Vector, What is a checkpoint
Choice, Score, Noul (TypeSafe): Apex, What gets bolted on
Classifier, regressor, clustering, anomaly detection: Nexus, Classical machine learning
CLIP, SigLIP, ViT, contrastive: Nexus, Vision understanding
Cloud: Origin, Where do you meet one
Coding agent: Origin, Where do you meet one; Apex, What gets bolted on
ComfyUI: Nexus, Image generation
Computer use, browser agent: Apex, What gets bolted on
Confabulation: Origin, How did it learn
Context engineering: Apex, What gets bolted on
Context window, context length: Vector, What is a token; Nexus, Language models: which variant
ControlNet: Nexus, Image generation
Conversational speech model: Nexus, Speech and audio
Cosine similarity, dimensions, Matryoshka (MRL): Nexus, Embeddings and rerankers
CPU, processor: Origin, What does it run on
CUDA: Origin, What does it run on

Decision model, System One (emerging): Apex, What gets bolted on
Decode, prefill, time to first token, tokens per second: Nexus, Language models: the dials
Deep learning, neural network, layer: Origin, What does "AI" actually mean on the box
Deepfake: Apex, How is it evaluated
Dense: Vector, What does "235B-A22B" mean
Diarisation, VAD: Nexus, Speech and audio
Diffusion, DiT, flow matching, rectified flow: Nexus, Image generation; Apex, What do the architecture words on a release post mean
Distillation: Apex, How was it made
DPO, GRPO, RLVR, RLHF, RLAIF: Apex, How was it made
Drift, regression, decay: Apex, How is it evaluated

Effective parameters: Nexus, Language models: which variant
Embedding (the layer): Vector, What is a parameter; (the model): Nexus, Embeddings and rerankers
Eval, eval harness, LLM-as-judge: Apex, How is it evaluated
Everyday AI, spam filter, translation: Origin, The AI you met today

Fine-tuning: Vector, What is a checkpoint; Apex, What did someone do to the base
Foundation model, frontier model: Vector, What does "multimodal" mean
Front, front end: Origin, Where do you meet one; Vector, What is in the folder
Function calling, tool use: Apex, What gets bolted on

Gated download: Vector, What is in the folder
Gateway, router, OpenRouter: Apex, What are the words on the invoice
Generative, generative AI: Origin, What does "AI" actually mean on the box
GGUF: Vector, What is in the folder
GPT, generative pre-trained transformer: Origin, What does "AI" actually mean on the box
GPU, graphics card: Origin, What does it run on
Graph engineering, loop engineering: Apex, What gets bolted on
GraphRAG, knowledge graph, semantic layer: Apex, What gets bolted on
Grounding: Nexus, Vision understanding; Apex, How is it evaluated
Guard models (Llama Guard, Prompt Guard): Apex, What gets bolted on

Hallucination: Origin, How did it learn
Harness (agent, eval): Apex, What gets bolted on; How is it evaluated
Hosted: Vector, What is inference
Hybrid search, semantic search: Nexus, Embeddings and rerankers

Inference: Origin, What does it do when it runs; Vector, What is inference
Inpainting, outpainting, image-to-image: Nexus, Image generation
Input tokens, output tokens, thinking tokens, cached tokens: Vector, What is a token

Jailbreak, red-teaming, sycophancy: Apex, How is it evaluated
JAX, PyTorch, TensorFlow: Vector, What is in the folder; Apex, Whose names are these
JEPA: Apex, What is a world model

Kernel: Vector, Why does one model have thirty downloads
Knowledge cutoff: Origin, How did it learn; Vector, What is in the folder
KV cache: Nexus, Language models: which variant

Latency, throughput: Vector, What is inference
Latent MoE, shared experts, fine-grained experts: Apex, What do the architecture words on a release post mean
Licence (Apache, MIT, CC, community, non-commercial): Vector, Are open weights open source; every Nexus licence trap
LLM, large language model: Origin, What does "AI" actually mean on the box; What does it do when it runs; Vector, What is a parameter
Local, on-device: Origin, Where do you meet one; Vector, What is inference
Logits, probability distribution, greedy decoding, seed, determinism: Nexus, Language models: the dials
Loop: Origin, What does it do when it runs
LoRA: Vector, What is a checkpoint

Machine learning, training data: Origin, What does "AI" actually mean on the box
Mamba, state-space, Mixture-of-Transformers: Apex, What do the architecture words on a release post mean
MCP: Apex, What gets bolted on
Memory (an agent's, a product's): Origin, How did it learn; Apex, What gets bolted on
Memory floor: every Nexus family section
Merge: Apex, What did someone do to the base
MMLU, MMLU-Pro, GPQA, SWE-bench: Apex, How do you read the table
mmproj, projector: Vector, Why does one model have thirty downloads; What does "multimodal" mean
Modality, multimodal, omni: Vector, What does "multimodal" mean
Model (the plain sense): Intro, What is "artificial intelligence"; Origin, Why is it a download
Model card, system card: Vector, What is in the folder; Apex, How is it evaluated
MoE, experts, routing, top-k: Vector, What does "235B-A22B" mean
MTEB: Nexus, Embeddings and rerankers
MTP, multi-token prediction: Nexus, Language models: the dials; Apex, What do the architecture words on a release post mean

Name fragments (date stamps, patch16, so400m, hiera, TDT, 12Hz, n-gram embedding, Engram, effective): Vector, What does Qwen3.8-27B-UD-Q4_K_M.gguf actually say
Narrow AI, general AI, AGI: Intro, What is "artificial intelligence"; Vector, What does "multimodal" mean
NPU: Origin, What does it run on

OCR, document understanding: Nexus, Vision understanding
Offloading: Origin, What does it run on; Nexus, Video generation
Ollama, LM Studio, llama.cpp, vLLM, SGLang: Nexus, Language models: which variant
Open weight, open source, OSAID: Vector, Are open weights open source
OpenAI-compatible endpoint: Nexus, Language models: which variant

Parameter, weight, tensor: Vector, What is a parameter
Post-training, pre-training, SFT: Apex, How was it made
Prompt, completion: Origin, What is it actually reading; Vector, What is inference
Prompt caching, semantic caching: Apex, What are the words on the invoice
Prompt injection: Apex, What gets bolted on; How is it evaluated
Pruning: Apex, How was it made

QAT, quantisation-aware training: Apex, How was it made
QLoRA: Vector, What is a checkpoint
Quantisation, Q4_K_M and the tag zoo: Vector, Why does one model have thirty downloads

RAG, chunking, vector database: Apex, What gets bolted on
Rate limits: Apex, What are the words on the invoice
Reasoning model, thinking budget, test-time compute: Vector, What is a thinking model; Nexus, Language models: which variant; Apex, What gets bolted on
Reranker, bi-encoder, cross-encoder: Nexus, Embeddings and rerankers
RLCD, reinforcement learning for calibrated decisions (reported): Apex, How was it made; What gets bolted on
RoPE, context extension: Apex, What do the architecture words on a release post mean
Runtime, engine, backend: Vector, What is inference

safetensors: Vector, What is in the folder
Segmentation, detection, classification: Nexus, Vision understanding
Self-hosted, provider-hosted: Origin, Where do you meet one; Vector, What is inference
Self-hosting versus API: Apex, What are the words on the invoice
Serving: Vector, What is inference
Small language model, SLM: Nexus, Language models: which variant
Spectrogram: Nexus, Speech and audio
Speculative decoding, draft model: Nexus, Language models: the dials
Step: Nexus, Image generation (same word, different thing)
Streaming: Apex, What are the words on the invoice
Structured outputs, JSON mode: Apex, What are the words on the invoice
Supervised, unsupervised, self-supervised: Apex, How was it made
Synthetic data: Apex, How was it made
System instructions, developer instructions: Vector, What is in the folder

Temperature, top-p, min_p, repetition penalty: Nexus, Language models: the dials
Text encoder, VAE: Nexus, Image generation
Thinking model, reasoning model: Vector, What is a thinking model
Token: Origin, What is it actually reading; Vector, What is a token
Tokenizer, vocabulary, BPE: Vector, What is a token
Training: Origin, How did it learn; Apex, How was it made
Transformer: Apex, What do the architecture words on a release post mean
TTS, voice cloning, vocoder: Nexus, Speech and audio

Uncensored: Apex, What did someone do to the base
Unified memory: Origin, What does it run on
Upscaler: Nexus, Image generation

VAE: Nexus, Image generation
Vector: Vector, Which words mean two things; Nexus, Embeddings and rerankers
Vibe coding: Apex, What gets bolted on
VLA, robot policy: Apex, What is a world model
VLM: Nexus, Vision understanding
VRAM: Origin, What does it run on

World model: Apex, What is a world model

Zero-shot: Nexus, Vision understanding

The Sources

IntroThe Intro

J. McCarthy, M. L. Minsky, N. Rochester and C. E. Shannon, "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence", dated 31 August 1955, as hosted on John McCarthy's Stanford pages (the phrase's first appearance, and the conjecture the intro paraphrases): http://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html

OriginTier 01

Qwen3.8-27B repository (file listing, config.json, model.safetensors.index.json, README.md, LICENSE), Hugging Face Hub, read via the Hub API: https://huggingface.co/Qwen/Qwen3.8-27B
Token counts in "What is it actually reading?" were produced by running the tokenizer.json from that repository with the Hugging Face tokenizers library on the sentences shown; the Urdu and Arabic sentences are my own renderings of the English one
OpenAI, "What are tokens and how to count them?" (rules of thumb; input, output, cached and reasoning token categories): https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them

VectorTier 02

PyTorch Foundation, "PyTorch Foundation Announces Safetensors as Newest Contributed Project to Secure AI Model Execution" (Paris, 8 April 2026; safetensors joins the Foundation as a hosted project): https://pytorch.org/blog/pytorch-foundation-announces-safetensors-as-newest-contributed-project-to-secure-ai-model-execution/
Qwen3.8-27B repository as above; tensor shapes read from the safetensors headers of shards 1 and 3; special tokens from tokenizer_config.json; template length from chat_template.jinja
Unsloth GGUF quantisations of Qwen3.8-27B (file list and sizes, licence, download count), read via the Hub API: https://huggingface.co/unsloth/Qwen3.8-27B-GGUF
Qwen3.8-Flash-Next model card, benchmark table, read 24 September 2026 (Qwen3.8-27B at 27B against Qwen3.7-Plus at 397B: the 27B scores above it on DeepSWE 1.1, SWE-bench Pro, CoWorkBench, JobBench, IFBench and LiveCodeBench v6, and below it on SWE-bench Multilingual and GPQA Diamond): https://huggingface.co/Qwen/Qwen3.8-Flash-Next
Fragments box at 2.1, read 24 September 2026: Qwen-Image-2512 card ("the December update"): https://huggingface.co/Qwen/Qwen-Image-2512
Fragments box at 2.1, read 24 September 2026: Mistral's models overview (Mistral Small 4 listed as v26.03, API name mistral-small-2603): https://docs.mistral.ai/getting-started/models/models_overview
Fragments box at 2.1, read 24 September 2026: DeepSeek-V4-Pro-0813 card (the suffix is not explained there): https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813
Fragments box at 2.1, read 24 September 2026: google/vit-base-patch16-224 card ("a sequence of fixed-size patches (resolution 16x16)"): https://huggingface.co/google/vit-base-patch16-224
Fragments box at 2.1, read 24 September 2026: Alabdulmohsin et al., "Getting ViT in Shape" (SoViT-400m, the shape-optimised ViT): https://arxiv.org/abs/2305.13035
Fragments box at 2.1, read 24 September 2026: facebookresearch/hiera ("Hiera: A Hierarchical Vision Transformer without the Bells-and-Whistles"): https://github.com/facebookresearch/hiera
Fragments box at 2.1, read 24 September 2026: facebookresearch/sam2 (the sam2.1_hiera checkpoints): https://github.com/facebookresearch/sam2
Fragments box at 2.1, read 24 September 2026: Xu et al., "Efficient Sequence Transduction by Jointly Predicting Tokens and Durations" (the Token-and-Duration Transducer): https://arxiv.org/abs/2304.06795
Fragments box at 2.1, read 24 September 2026: Qwen3-TTS-12Hz-1.7B-Base card (its table lists the 12Hz tokenizer at 12.5 frames per second): https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-Base
Fragments box at 2.1, read 24 September 2026: Qwen3.8-Flash-Next card (the n-gram embedding: "indexing with short n-grams", "more amenable to offloading than Mixture-of-Experts"): https://huggingface.co/Qwen/Qwen3.8-Flash-Next
Fragments box at 2.1, read 24 September 2026: DeepSeek-V4.1-Flash card ("Engram conditional memory (196B parameters, sparsely accessed via token-based lookup)"; 8B active in prefill, 16B in decode): https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
Fragments box at 2.1, read 24 September 2026: gemma-4-E4B-it card ("4.5B effective (8B with embeddings)", per-layer embeddings): https://huggingface.co/google/gemma-4-E4B-it
Qwen3.8-27B file listing via the Hub API, read 24 September 2026 (crc32.txt; an LFS oid hash on every large file): https://huggingface.co/api/models/Qwen/Qwen3.8-27B/tree/main
Hugging Face Hub, model cards documentation and the model card metadata specification, read 24 September 2026 (no cutoff key among the metadata): https://huggingface.co/docs/hub/model-cards and https://raw.githubusercontent.com/huggingface/hub-docs/main/modelcard.md
Claude API documentation, context windows, read 24 September 2026 ("If the input alone already exceeds the model's context window, the API returns a 400 invalid_request_error"): https://platform.claude.com/docs/en/build-with-claude/context-windows
llama.cpp server README, read 24 September 2026 (a truncated flag when prompt plus generated tokens exceed n_ctx; n_keep for what to retain when tokens are discarded): https://raw.githubusercontent.com/ggml-org/llama.cpp/master/tools/server/README.md
Anthropic, Commercial Terms of Service ("Customer (a) retains all rights to its Inputs, and (b) owns its Outputs"; "may not train models on Customer Content") and its privacy centre on retention, read 24 September 2026, as one example of the words a provider's terms use, not as a statement of any provider's policy: https://www.anthropic.com/legal/commercial-terms and https://privacy.claude.com/en/articles/7996866-how-long-do-you-store-my-data
Ollama library page for qwen3.8:27b (default build: 18 GB, Q4_K_M model at 17 GB, 461M-parameter BF16 projector at 931 MB, 27.3B parameters): https://ollama.com/library/qwen3.8:27b
Qwen3-235B-A22B config (128 experts, 8 per token): https://huggingface.co/Qwen/Qwen3-235B-A22B
Gemma 4 26B-A4B config (128 experts, top-8): https://huggingface.co/google/gemma-4-26B-A4B-it
Wan 2.2 T2V-A14B card metadata: https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B
Licence and gating fields read via the Hub API on 20 September 2026 for: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct , https://huggingface.co/black-forest-labs/FLUX.1-dev , https://huggingface.co/black-forest-labs/FLUX.2-dev , https://huggingface.co/black-forest-labs/FLUX.1-schnell , https://huggingface.co/black-forest-labs/FLUX.2-klein-4B , https://huggingface.co/stabilityai/stable-diffusion-3.5-large , https://huggingface.co/mlx-community/Qwen3.8-27B-4bit , https://huggingface.co/mlx-community/Qwen3.8-27B-8bit
Hugging Face, safetensors repository and format specification (8-byte header length, JSON header, no code execution; pickle background): https://github.com/huggingface/safetensors
DataCamp, safetensors joining the PyTorch Foundation in April 2026 (secondary; the move is reported here, not verified against a foundation announcement): https://www.datacamp.com/blog/safetensors-format
GGUF specification (single-file format, metadata plus tensors, successor to GGML/GGMF/GGJT): https://github.com/ggerganov/ggml/blob/9c2adc4962a3a5d259f10db2171e0df5c83e4b05/docs/gguf.md and the Hugging Face Transformers GGUF page: https://huggingface.co/docs/transformers/en/quantization/gguf.md
GGUF introduction date, August 2023 (Wikipedia, secondary; the llama.cpp pull request that merged the spec is dated November 2023): https://en.wikipedia.org/wiki/GGUF and https://github.com/ggml-org/ggml/pull/302
Hugging Face Hub documentation, gated models (automatic and manual approval): https://huggingface.co/docs/hub/models-gated
Hu et al., "LoRA: Low-Rank Adaptation of Large Language Models", arXiv 2106.09685 v2 (frozen weights, injected low-rank matrices, up to 10,000 times fewer trainable parameters on GPT-3): https://arxiv.org/abs/2106.09685v2
Dettmers et al., "QLoRA: Efficient Finetuning of Quantized LLMs" (frozen 4-bit base, LoRA adapters, 65B on a 48 GB GPU): https://arxiv.org/abs/2305.14314v1
Apache License 2.0 as shipped in the Qwen3.8-27B folder (clauses 4(b), 4(c), 4(d) and 6 for notices, modified files, NOTICE files and trademarks)
Open Source Initiative, The Open Source AI Definition 1.0 (data information, code, parameters; adopted 27 October 2024): https://opensource.org/ai/open-source-ai-definition and board minutes https://opensource.org/meeting-minutes/2024-10-27
Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni and Percy Liang, "Lost in the Middle: How Language Models Use Long Contexts" (Transactions of the Association for Computational Linguistics, volume 12, 2024, pages 157 to 173; arXiv preprint 2023; performance is often highest when relevant information is at the beginning or end of the input and significantly degrades when it is in the middle of a long context): https://arxiv.org/abs/2307.03172

NexusTier 03, language models (all read via the Hub API on 20 September 2026: licence and gating fields, safetensors parameter counts, `config.json` context and expert settings; card text where quoted)

https://huggingface.co/Qwen/Qwen3.8-27B (thinking mode, reasoning_effort, named runtimes)
https://huggingface.co/Qwen/Qwen3.8-Flash-Next
bartowski's GGUF quantisations of Qwen3.8-27B, card read 24 September 2026 (perplexity and KL divergence of each quantisation against the BF16 reference): https://huggingface.co/bartowski/Qwen3.8-27B-GGUF
llama.cpp, perplexity tool README, read 24 September 2026 (what perplexity measures; KL divergence of a quantised model's logits against an FP16 reference, 0 meaning identical): https://github.com/ggml-org/llama.cpp/blob/master/tools/perplexity/README.md
https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash and https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813
vLLM documentation, "Parallelism and Scaling", read 24 September 2026 (tensor parallelism "if the model is too large for a single GPU"; the GPU KV cache size and Maximum concurrency lines at startup): https://docs.vllm.ai/en/latest/serving/parallelism_scaling.html
Databricks, "LLM Inference Performance Engineering: Best Practices", read 24 September 2026 (decode speed set by loading parameters from memory; batching "increases throughput compared to processing queries sequentially, but each query will take longer to complete"; the cache grows with the number of sequences and their lengths): https://www.databricks.com/blog/llm-inference-performance-engineering-best-practices
Rouhani et al., "Microscaling Data Formats for Deep Learning", read 24 September 2026 (MX formats "combine a per-block scaling factor with narrow floating-point and integer types"): https://arxiv.org/abs/2310.10537
https://huggingface.co/moonshotai/Kimi-K3
https://huggingface.co/zai-org/GLM-5.3
https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct (17B activated, 10M context, early fusion, system protections) and https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct (17B activated, 1M context)
https://huggingface.co/google/gemma-4-26B-A4B-it (25.2B total, 3.8B active, 8 active of 128 plus 1 shared, 256K context per card) , https://huggingface.co/google/gemma-4-12B-it , https://huggingface.co/google/gemma-4-E4B-it , https://huggingface.co/google/gemma-4-E2B-it (effective-parameter naming, thinking token budgets, language counts, audio on E2B/E4B/12B)
https://huggingface.co/openai/gpt-oss-120b and https://huggingface.co/openai/gpt-oss-20b (reasoning effort, chain of thought not for end users, MXFP4 and the 80 GB / 16 GB floors)
https://huggingface.co/mistralai/Mistral-Small-4-119B-2603 (119B, 6.5B activated per token, branded A6B; 128 experts 4 active; 256k context per card; config exposes 1,048,576 positions)
https://huggingface.co/microsoft/Phi-4
https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
Ollama library entry for qwen3.8:27b (parameters object, draft count): https://ollama.com/library/qwen3.8:27b
LM Studio named as a GGUF tool on the secondary Wikipedia GGUF page cited under Tier 02

NexusTier 03, image generation (Hub API fields and card text, 20 September 2026)

https://huggingface.co/black-forest-labs/FLUX.2-dev (32B rectified flow transformer)
https://huggingface.co/black-forest-labs/FLUX.2-klein-4B (~13 GB VRAM, RTX 3090/4070 and above)
https://huggingface.co/stabilityai/stable-diffusion-3.5-large (MMDiT; three text encoders: OpenCLIP-ViT/G, CLIP-ViT/L, T5-xxl) and https://huggingface.co/stabilityai/stable-diffusion-3.5-medium
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
https://huggingface.co/Qwen/Qwen-Image-2512 (text rendering claim) and https://huggingface.co/Qwen/Qwen-Image
https://huggingface.co/Tongyi-MAI/Z-Image-Turbo (8 NFEs, sub-second claim)
https://huggingface.co/tencent/HunyuanImage-3.0 (64 experts, 80B total, 13B active)
https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px_diffusers

NexusTier 03, video generation (Hub API fields and card text, 20 September 2026)

https://huggingface.co/Lightricks/LTX-2.5 (audio modes, revenue-tiered licence) and https://huggingface.co/Lightricks/LTX-2
https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B (80 GB statement) and https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B (24 GB / RTX 4090 statement; 720p 24 fps; MoE in a video diffusion model)
https://huggingface.co/tencent/HunyuanVideo-1.5 (14 GB minimum with offloading)
https://huggingface.co/genmo/mochi-1-preview (22 / 42 / 60 GB statements)
https://huggingface.co/zai-org/CogVideoX-5b
https://huggingface.co/stepfun-ai/stepvideo-t2v

NexusTier 03, speech and audio (Hub API fields and card text, 20 September 2026)

NexusTier 03, vision understanding (Hub API fields and card text, 20 September 2026)

NexusTier 03, embeddings and rerankers (Hub API fields and card text, 20 September 2026)

https://huggingface.co/Qwen/Qwen3-Embedding-0.6B (dimensions 32 to 1,024; 100+ languages; MTEB claim), https://huggingface.co/Qwen/Qwen3-Embedding-4B , https://huggingface.co/Qwen/Qwen3-Embedding-8B , https://huggingface.co/Qwen/Qwen3-Reranker-0.6B , https://huggingface.co/Qwen/Qwen3-VL-Embedding-2B
https://huggingface.co/google/embeddinggemma-300m (768 dims, MRL 512/256/128, on-device)
https://huggingface.co/jinaai/jina-embeddings-v5-text-small and https://huggingface.co/jinaai/jina-reranker-v3
https://huggingface.co/BAAI/bge-m3 , https://huggingface.co/nomic-ai/nomic-embed-text-v2-moe , https://huggingface.co/vidore/colpali-v1.3 , https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2 , https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1
Muennighoff et al., "MTEB: Massive Text Embedding Benchmark" (8 tasks, 58 datasets, 112 languages; leaderboard address): https://arxiv.org/pdf/2210.07316v2 and the Hugging Face MTEB blog post https://huggingface.co/blog/mteb

NexusTier 03, classical machine learning

scikit-learn user guide, table of contents (supervised, unsupervised and semi-supervised sections; ensembles): https://scikit-learn.org/stable/user_guide

ApexTier 04, the decision-model box (all secondary or vendor; nothing verified against weights or a paper)

DataCamp, "Jev: TypeSafe's System One Model" (announcement date, category name, input and output shape, Jevons naming): https://www.datacamp.com/blog/system-one-models-jev
Runtime Wire, "TypeSafe opens Jev early access" (no weights or reproducible paper; RLCD named; Kahneman reference; speed claims vendor-tested): https://runtimewire.com/article/typesafe-jev-system-one-ai-model-early-access
Pasquale Pillitteri, "TypeSafe launches Jev" (Sean Goedecke's analysis: prefill plus single-token decoding, reproduction attempts, can still choose wrongly, capability ceiling): https://pasqualepillitteri.it/en/news/16460/typesafe-jev-chatless-ai-beats-claude
TypeSafe AI documentation, "Introduction" (the three question shapes and what each returns; questions over one state evaluated in parallel and in isolation; ask one well-scoped thing per question and combine in code): https://docs.typesafe.ai/introduction
TypeSafe AI documentation, "Confidence" (confidence as the concentration of the returned distribution; Noul answers carry none): https://docs.typesafe.ai/confidence
TypeSafe AI documentation, "System One" (trained for calibrated decisions; probabilities optimised against outcomes): https://docs.typesafe.ai/concepts/system-one

ApexTier 04, training

TypeSafe AI documentation, "Machine learning primer" (RLCD named; calibration as outcomes at 0.8 occurring about 80 per cent of the time; the contrast with RLHF): https://docs.typesafe.ai/introduction/machine-learning-primer
"Scaling Laws for Neural Language Models" (arXiv 2001.08361; the loss scales as a power law with model size, dataset size and training compute): https://arxiv.org/abs/2001.08361
DeepSeek-R1 card (R1-Zero trained by RL without SFT; cold-start data; six distilled models; 32B claim): https://huggingface.co/deepseek-ai/DeepSeek-R1 and https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
Ouyang et al., "Training language models to follow instructions with human feedback" (InstructGPT, RLHF): https://arxiv.org/abs/2203.02155
Bai et al., "Constitutional AI" (RLAIF): https://arxiv.org/abs/2212.08073
Rafailov et al., "Direct Preference Optimization": https://arxiv.org/abs/2305.18290
Shao et al., "DeepSeekMath" (GRPO): https://arxiv.org/abs/2402.03300
DeepSeek-AI, "DeepSeek-R1" paper: https://arxiv.org/abs/2501.12948
Phi-4 card (synthetic "textbook-like" data): https://huggingface.co/microsoft/phi-4
Minitron-8B-Base card (pruning of Nemotron-4 15B, then distillation): https://huggingface.co/nvidia/Minitron-8B-Base
Gemma 4 QAT repositories, e.g. https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-gguf and https://huggingface.co/google/gemma-4-26B-A4B-it-qat-q4_0-unquantized
scikit-learn user guide (supervised, unsupervised, semi-supervised sections): https://scikit-learn.org/stable/user_guide

ApexTier 04, architecture words

Vaswani et al., "Attention Is All You Need" (arXiv 1706.03762, June 2017; the Transformer, an architecture based on attention alone): https://arxiv.org/abs/1706.03762
Qwen3.8-27B card and config (hidden layout, 48 linear-attention and 16 full-attention layers, 24 query and 4 key-value heads, MTP): https://huggingface.co/Qwen/Qwen3.8-27B
Kimi-K3 card and config (2.8T total, 104B activated, 896 experts, 16 selected, 2 shared, 1,048,576 context; the card names its MoE "Stable LatentMoE"): https://huggingface.co/moonshotai/Kimi-K3
Latent MoE paper (NVIDIA, 2026): https://arxiv.org/html/2601.18089v1 ; Sebastian Raschka's architecture gallery entry (secondary): https://sebastianraschka.com/llm-architecture-gallery/latent-moe/
Cosmos 3 Super card (Mixture-of-Transformers, autoregressive and diffusion towers): https://huggingface.co/nvidia/Cosmos3-Super

ApexTier 04, variants

Abliterated Qwen3.8-27B repositories found by hub search on 20 September 2026, e.g. https://huggingface.co/huihui-ai/Huihui-Qwen3.8-27B-abliterated (Apache 2.0)
Abliterated Llama 3.1 example carrying the llama3.1 licence field: https://huggingface.co/mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated
Guard models: https://huggingface.co/meta-llama/Llama-Guard-4-12B and https://huggingface.co/meta-llama/Llama-Prompt-Guard-2-86M ; the Llama 4 Scout card's system-protections statement

ApexTier 04, bolt-ons and agents

Anthropic, "Donating the Model Context Protocol and establishing the Agentic AI Foundation" (9 December 2025; co-founders; 10,000+ servers): https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
OWASP, LLM01:2025 Prompt Injection (direct and indirect injection; a modified document altering an application's output), read 24 September 2026: https://genai.owasp.org/llmrisk/llm01-prompt-injection/
Linux Foundation, launch of the Agent2Agent protocol project (23 June 2025; protocol created by Google): https://s24.q4cdn.com/538403808/files/doc_news/Linux-Foundation-Launches-the-Agent2Agent-Protocol-Project-to-Enable-Secure-Intelligent-Communication-Between-AI-Agents-2025.pdf
AAIF founding projects including AGENTS.md (secondary): https://agentic-ai.readthedocs.io/en/latest/Standards/agentic-ai-foundation/
Engineering-ladder attestations (all secondary; usage evidence only): https://medium.com/@neuraldev/loop-engineering-vs-graph-engineering-the-architecture-shift-quietly-reshaping-ai-agents-c83488435d23 , https://www.analyticsvidhya.com/blog/2026/08/agent-harness-loop-graph-engineering/ , https://towardsdatascience.com/graph-engineering-for-ai-agents-from-prompts-and-loops-to-workflows/ , https://github.com/ai-boost/awesome-harness-engineering , https://www.sitepoint.com/vibe-coding-2026-complete-guide/

ApexTier 04, words on the invoice

OpenAI, prompt caching guide, read 20 September 2026 (whole rendered prefix must match; minimum cacheable length, read and write ratios and breakpoint controls are per model generation; cached_tokens reporting): https://developers.openai.com/api/docs/guides/prompt-caching.md
OpenAI, batch endpoint reference (24-hour completion window is the only one supported; 50,000 requests per file): https://developers.openai.com/api/reference/cli/resources/batches/methods/create/index.md
OpenAI, structured outputs guide (function calling or json_schema response format): https://developers.openai.com/api/docs/guides/structured-outputs
OpenRouter, structured outputs documentation (gateway example; per-model parameter support): https://openrouter.ai/docs/guides/features/structured-outputs

ApexTier 04, world models, robotics, 3D (Hub API fields and card text, 20 September 2026)

ApexTier 04, evaluation and provenance

MTEB paper, "no particular text embedding method dominates across all tasks": https://arxiv.org/pdf/2210.07316v2
EleutherAI, lm-evaluation-harness ("a unified framework to test generative language models"; "the backend for Hugging Face's popular Open LLM Leaderboard"), read 24 September 2026: https://github.com/EleutherAI/lm-evaluation-harness
Evaluation glossary showing the regression/drift/decay disagreement (secondary): https://futureagi.com/blog/llm-evaluation-glossary-definitions-2026/
Google Cloud documentation, Content Credentials (C2PA description; validation failure on non-C2PA modification): https://docs.cloud.google.com/vertex-ai/generative-ai/docs/content-credentials
European Commission, Code of Practice on Transparency of AI-Generated Content: press release of 10 June 2026 (obligations apply from 2 August 2026) https://digital-strategy.ec.europa.eu/en/news/commission-publishes-code-practice-marking-and-labelling-ai-generated-content and FAQ https://digital-strategy.ec.europa.eu/en/faqs/signing-code-practice-transparency-ai-generated-content
Google DeepMind, SynthID technology page (watermarking and identification across images, audio, text and video; beta): https://deepmind.google/technologies/synthid/
Further SynthID and provenance reading (secondary): https://compliancehub.wiki/eu-ai-act-marking-labelling-code-of-practice-article-50-2026/ , https://c2paviewer.com/articles/eu-ai-content-labels-c2pa , https://c2paviewer.com/articles/verify-ai-generated-image-c2pa-synthid , https://mvidmar.substack.com/p/ai-watermark-illusion-en

ApexTier 04, reading the table

"Measuring Massive Multitask Language Understanding" (MMLU; arXiv 2009.03300; 57 tasks from elementary mathematics to law): https://arxiv.org/abs/2009.03300
"MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark" (arXiv 2406.01574; trivial questions removed, ten options): https://arxiv.org/abs/2406.01574
"GPQA: A Graduate-Level Google-Proof Q&A Benchmark" (arXiv 2311.12022; 448 questions written by domain experts in biology, physics and chemistry): https://arxiv.org/abs/2311.12022
"SWE-bench: Can Language Models Resolve Real-World GitHub Issues?" (arXiv 2310.06770; edit a codebase to resolve a described issue): https://arxiv.org/abs/2310.06770
"Humanity's Last Exam" (arXiv 2501.14249; 2,500 questions across dozens of subjects, written by subject-matter experts): https://arxiv.org/abs/2501.14249
"Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference" (arXiv 2403.04132; crowdsourced pairwise comparison): https://arxiv.org/abs/2403.04132

Corrections

None yet. Corrections are logged here with the date and the source that prompted them.