Agent Atlas

Reference · read 14 November 2026

The Model Shelf

Some models you can only rent. Others you can download and run yourself, if your machine has the memory. Here is what exists, where to get it, and what it takes to run.

Closed

Rented by the token

The frontier labs’ best models. You reach them through an app, a subscription or an API. The weights never leave the lab.

Open, large

Downloadable, heavy

Weights anyone can download. Running the biggest takes a server; most people use them through a cloud host.

Open, small

Runs at home

A few billion parameters. A laptop runs them offline, privately, for free, and less capably.

Frontier models, closed

Factual
Anthropic, OpenAI, Google and xAI. The strongest models, rented, never downloaded.

These are what coding agents use. Their size is not published. The Frontier snapshot has context windows and knowledge cutoffs for eight labs, dated.

Claude Fable 5.2

Anthropic

Anthropic's top model; $8 in / $40 out per million tokens. Anthropic recommends Opus 5.5 ($4 / $20) as the default.

Context
1M

(Anthropic)

Gemini 4 Carbon

Google

Google's strongest reasoning model; price doubles on 1 Dec 2026 ($2→$4 in, $10→$20 out), see google-gemini-pricing.

Context
1,048,576

(Google)

GPT-6 Astra

OpenAI

OpenAI's named flagship; $10 in / $50 out per million tokens.

Context
1.05M

(OpenAI)

GPT-6 Luna

OpenAI

Cheapest tier; $0.1 in / $0.5 out per million tokens.

Context
1.05M

(OpenAI)

Grok 4.7

xAI

xAI's flagship and only listed text model; $2.00 in / $6.00 out per million tokens.

Context
500k

(xAI)

Small open models

Factual
Under about 15 billion parameters. They run on a laptop, a mini PC, some on a phone.

SmolLM3-3B

Hugging Face

Fully open including training data recipe. Ollama's smollm page is the older SmolLM, not SmolLM3.

Size
3B
License
Apache 2.0
Context
128k
Runs on
This site’s M1 miniabout 1.7 GB at 4-bit

Hugging Face(Hugging Face)

Qwen3.5-4B

Alibaba Qwen

Laptop or high-end phone; 3.3-4.0GB on Ollama. 0.8B and 2B sizes also exist.

Size
4B
License
Apache 2.0
Context
262,144
Runs on
This site’s M1 miniabout 2.3 GB at 4-bit

Hugging FaceOllama(Alibaba Qwen)

Gemma 4 E2B

Google DeepMind

Smallest Gemma 4, for phones; the model Ollama's quickstart pulls first.

Size
2.3B effective (5.1B with embeddings)
License
Apache 2.0
Context
128K
Runs on
This site’s M1 miniabout 2.9 GB at 4-bit

Hugging FaceOllama(Google DeepMind)

Gemma 4 E4B

Google DeepMind

Built for phones and edge devices; text, image and audio input.

Size
4.5B effective (8B with embeddings)
License
Apache 2.0
Context
128K
Runs on
This site’s M1 miniabout 4.6 GB at 4-bit

Hugging FaceOllama(Google DeepMind)

Nemotron 3.5 Bolt E4B

NVIDIA

NVIDIA's retune of Gemma 4 E4B for local voice and device assistants; 7.2-7.9GB on Ollama.

Size
8B, 4B effective
License
NVIDIA OpenMDW License 1.1
Context
128K
Runs on
This site’s M1 miniabout 4.6 GB at 4-bit

Hugging FaceOllama(NVIDIA)

Qwen3.5-9B

Alibaba Qwen

Newest small Qwen (Qwen3.8 has no small sizes); 6.6-7.6GB on Ollama.

Size
9B
License
Apache 2.0
Context
262,144
Runs on
This site’s M1 miniabout 5.1 GB at 4-bit

Hugging FaceOllama(Alibaba Qwen)

gpt-oss-20b

OpenAI

Over 15B total but only 3.6B active; runs within 16GB of memory, 14GB on Ollama.

Size
21B total, 3.6B active
License
Apache 2.0
Context
128K
Runs on
A base Mac mini or MacBookabout 12 GB at 4-bit

Hugging FaceOllama(OpenAI)

Large open models

Factual
Free to download, but most need a workstation or a server to run.

Qwen3.8-27B

Alibaba Qwen

Single-GPU vision-language model; 18GB download on Ollama. Qwen's 2.4T flagship weights are also open (Qwen3.8-2.4T-A95B).

Size
27B (dense)
License
Apache 2.0
Context
262,144
Runs on
A 32 GB mini PC or Macabout 15 GB at 4-bit

Hugging FaceOllama(Alibaba Qwen)

Gemma 4 31B

Google DeepMind

Biggest Gemma 4, aimed at consumer GPUs and workstations; 19-20GB on Ollama. A 26B A4B MoE sibling has 3.8B active.

Size
30.7B (dense)
License
Apache 2.0
Context
256K
Runs on
A 32 GB mini PC or Macabout 17 GB at 4-bit

Hugging FaceOllama(Google DeepMind)

Llama 4 Scout

Meta

Meta's newest open weights are still Llama 4 (Apr 2025). Fits one H100 at int4; Maverick sibling is 400B total.

Size
109B total, 17B active
License
Llama 4 Community License
Context
10M
Runs on
DGX Spark, or a big Mac Studioabout 62 GB at 4-bit

Hugging FaceOllama(Meta)

gpt-oss-120b

OpenAI

Fits one 80GB GPU thanks to 4-bit MXFP4 weights; 65GB on Ollama.

Size
117B total, 5.1B active
License
Apache 2.0
Context
128K
Runs on
DGX Spark, or a big Mac Studioabout 67 GB at 4-bit

Hugging FaceOllama(OpenAI)

Nemotron 3.5 Super

NVIDIA

Vision-language model for coding and tools; NVIDIA serves it on 4× H100; 80-90GB on Ollama.

Size
120B total, 12B active
License
NVIDIA OpenMDW License 1.1
Context
1M
Runs on
DGX Spark, or a big Mac Studioabout 68 GB at 4-bit

Hugging FaceOllama(NVIDIA)

GLM-5.3-Flash

Z.ai (Zhipu)

Multimodal coding and agent model. The bigger GLM-5.3 (753B) uses a custom license. Ollama cloud only.

Size
320B total, 18B active
License
MIT
Context
1M
Runs on
A server with eight H100sabout 182 GB at 4-bit

Hugging FaceOllama(Z.ai)

MiniMax-M3

MiniMax

Multimodal coding and agent model with sparse attention for long context. Ollama cloud only, served at 512K.

Size
~428B total, ~23B active
License
MiniMax Community License
Context
1M
Runs on
A server with eight H100sabout 244 GB at 4-bit

Hugging FaceOllama(MiniMax)

DeepSeek-V4.1-Flash

DeepSeek

Multimodal MoE built for cheap long agent runs. A 1017 update exists. Ollama cloud only.

Size
552B backbone (763B in files), 16B active when generating
License
MIT
Context
1M
Runs on
A server with eight H100sabout 315 GB at 4-bit

Hugging FaceOllama(DeepSeek)

Mistral Large 4

Mistral AI

Largest European open-weight model; selling it as a hosted service needs a separate license. Ollama cloud only.

Size
1T total, 49B active (52B with embeddings)
License
Mistral Ecosystem Open-Weight License (modified MIT)
Context
1M
Runs on
Several serversabout 570 GB at 4-bit

Hugging FaceOllama(Mistral AI)

Kimi K3.1 Pro

Moonshot AI

Largest open-weight model listed; datacenter only. Ollama offers it as cloud only.

Size
2.8T total, 104B active
License
Kimi K3.1 Pro License (custom)
Context
1048576
Runs on
Several serversabout 1,595 GB at 4-bit

Hugging FaceOllama(Moonshot AI)

What machine runs what

Estimate
Every open model on one scale of memory, against the machines that hold it.

A model has to fit in memory to run at a usable speed. At 4-bit, the common download, each billion parameters takes about 0.57 GB, from llama.cpp’s measured 4.8944 bits per weight (ggml-org). We call a model a fit when its weights take under three quarters of the memory, leaving room for the conversation and the system. Mixture-of-experts models only compute with some of their weights per word, but all of them have to be loaded.

Log scale: each step to the right is ten times more memory. Dashed lines are three quarters of each machine’s memory. Our arithmetic on the published sizes.

Try any size and precision in Will it fit? on Plate V. For agents, not models, see how much machine you need.

Where to get them

Factual
Hugging Face hosts the weights. Ollama and LM Studio download and run them in one step.

Hugging Face

Where the weights live. Each model is a versioned repository of files, with a model card from its maker (Hugging Face). Every “Hugging Face” link above goes to one.

Ollama

Downloads and runs a model with one command, locally or in its cloud (Ollama). Some of the largest are offered there only in the cloud; the card says so.

ollama run gpt-oss:20b

LM Studio

A desktop app for macOS, Windows and Linux that finds, downloads and chats with models from Hugging Face (LM Studio). The easiest start if you do not want a terminal.

Start small. A 4 to 9 billion parameter model on a 16 GB machine answers quickly and teaches you what local models are good at, and where the frontier still wins.

Abliterated models

Factual
Open models with the refusal behavior cut out of the weights. What that means and what it costs.

Chat models are trained to refuse some requests. In 2024, researchers found that across 13 open chat models that refusal runs through a single direction inside the model. Erase it and the model stops refusing; add it and the model refuses even harmless requests (arXiv (Arditi, Obeso, Syed, Paleka, Panickssery, Gurnee, Nanda)).

Abliteration makes that permanent. Someone finds the refusal direction, then edits the weights so no layer can write along it, with no retraining (Maxime Labonne (Hugging Face community blog)). The result is posted as a new model, usually with “abliterated” in its name.

It has costs. In Maxime Labonne’s own test it degraded the model’s quality, and extra training won back most of the loss (Maxime Labonne (Hugging Face community blog)). It also removes refusals across the board, the reasonable ones included, and says nothing about whether the answers are true.

An uncensored model is a different route to a similar place: its maker deletes refusals from the training data and fine-tunes again (Eric Hartford).

Why it matters to everyone else: both papers conclude that safety training on open weights is easy to undo (arXiv (Arditi, Obeso, Syed, Paleka, Panickssery, Gurnee, Nanda); Maxime Labonne (Hugging Face community blog)). Once weights are public, the maker’s guardrails are a default, not a lock. We do not list abliterated downloads here.