Field guide · updated for July 2026

Open-weight models,
built by the giants.

A working comparison of large-company-backed models you can download and run entirely offline via Ollama — what they cost to use, how much RAM each size needs, and who actually keeps shipping updates.

Companies covered7

Model families10

All pullable via Ollama10 of 10

Apache 2.0 / MIT6 of 10

Open license — few or no strings Permissive with conditions Custom / restrictive license

Filter

Showing 10 of 10 models

On RAM figures: based on the default Q4_K_M quantization Ollama pulls by default — the realistic "what you'll actually download" number, not the full-precision size. Add roughly 10–20% on top for context/KV cache at longer conversations. GPU VRAM works the same way if you're running on a graphics card instead of system RAM.

On the "roughly comparable to" column: now broken down per size, since a 1B model and a 405B model from the same family are not in the same league. These are capability-tier approximations drawn from public benchmarks (MMLU, coding, reasoning suites) and vendor statements — not exact 1:1 matches. Note: DeepSeek's smaller tags (1.5B–70B) are Qwen/Llama models distilled on R1's reasoning traces, not the full 671B model — treat their equivalence loosely.

Model	Developer	Sizes & RAM needed (Ollama, Q4)	License	Commercial use	Roughly comparable to	Best for
Granite 4.1 IBM · `ollama pull granite4`	IBM	1B — ~2 GB RAM 3B — ~4 GB RAM 8B — ~6 GB RAM 30B — ~20 GB RAM	Apache 2.0	UnrestrictedNo usage cap, no MAU threshold	1B — below GPT-3.5 3B — ≈ GPT-3.5-Turbo 8B — ≈ GPT-4o mini 30B — ≈ GPT-4o mini+	Enterprise/regulated use, long context (512K), ISO 42001 compliance
Llama 3.3 / 4 Meta · `ollama pull llama3.3`	Meta	1B — ~2 GB RAM 8B — ~6 GB RAM 70B — ~43 GB RAM 405B — ~231 GB RAM	Llama License	ConditionalFree under 700M monthly active users; license required above	1B — below GPT-3.5 8B — ≈ GPT-3.5-Turbo 70B — ≈ GPT-4o mini 405B — ≈ GPT-4o	Widest tooling/ecosystem support, general-purpose all-rounder
Gemma 3 / 4 Google · `ollama pull gemma3`	Google	1B — ~1.5 GB RAM 4B — ~4.2 GB RAM 12B — ~8 GB RAM 27B — ~17 GB RAM	Gemma License	Mostly freePermissive but not OSI-recognized; bans certain use categories (weapons, surveillance)	1B — below GPT-3.5 4B — ≈ GPT-3.5-Turbo 12B — ≈ GPT-3.5-Turbo+ 27B — ≈ GPT-4o mini	Edge/mobile deployment, multimodal (image understanding)
Phi-4 Microsoft · `ollama pull phi4`	Microsoft	Mini (3.8B) — ~2.5 GB RAM Medium (14B) — ~9 GB RAM	MIT	UnrestrictedFully permissive MIT license	Mini (3.8B) — ≈ GPT-3.5-Turbo Medium (14B) — ≈ GPT-4o mini	Best reasoning-per-parameter on very low RAM (2–4GB)
Qwen 2.5 / 3.5 Alibaba · `ollama pull qwen2.5`	Alibaba	0.5B — ~1 GB RAM 7B — ~5.5 GB RAM 32B — ~20 GB RAM 72B+ — ~45 GB+ RAM	Apache 2.0	UnrestrictedNo royalties, no usage limits, redistribution allowed	0.5B — below GPT-3.5 7B — ≈ GPT-3.5-Turbo 32B — ≈ GPT-4o mini 72B+ — ≈ GPT-4o	Multilingual (200+ languages), strongest raw coding at full size
Mistral / Mixtral Mistral AI · `ollama pull mistral`	Mistral AI	7B — ~4.1 GB RAM 8x7B — ~26 GB RAM 8x22B — ~80 GB RAM	Apache 2.0	UnrestrictedShifted from earlier restrictive terms — now fully Apache 2.0	7B — ≈ GPT-3.5-Turbo 8x7B — ≈ GPT-4o mini 8x22B — ≈ GPT-4-Turbo	Long-document efficiency (sliding window attention), function calling
gpt-oss OpenAI · `ollama pull gpt-oss`	OpenAI	20B — ~16 GB RAM 120B — ~80 GB RAM (single GPU)	Apache 2.0	UnrestrictedSubject to OpenAI's gpt-oss usage policy	20B — ≈ o3-mini 120B — ≈ o4-mini	Near o4-mini reasoning locally; 20B runs on 16GB edge devices
DBRX Databricks · `ollama pull dbrx`	Databricks	132B (36B active) — ~80 GB RAM	Databricks OML	ConditionalFree under 700M MAU, similar structure to Llama license	132B — ≈ GPT-4o mini	Enterprise data workflows tied to Databricks stack
DeepSeek-R1 / V3 DeepSeek AI · `ollama pull deepseek-r1`	DeepSeek	1.5B — ~2 GB RAM 8B — ~6 GB RAM 32B — ~20 GB RAM 70B — ~43 GB RAM 671B — ~404 GB RAM	MIT	UnrestrictedFully permissive; note: some orgs restrict use on procurement/origin grounds	1.5B — below GPT-3.5 8B — ≈ GPT-3.5-Turbo 32B — ≈ o3-mini 70B — ≈ o3-mini+ 671B — ≈ o1	Frontier-level reasoning comparable to o1, fully open
Command A Cohere · `ollama pull command-a`	Cohere	111B — ~67 GB RAM Command R7B — ~6 GB RAM	CC-BY-NC	Non-commercialBlocked for commercial deployment without separate license	111B — ≈ GPT-4o mini R7B — ≈ GPT-3.5-Turbo	RAG and grounding-heavy workloads, research/non-commercial use

Notably absent

the labs everyone asks about first

Anthropic — Claude

No open-weight Claude models exist. Every Claude model is served exclusively via API or the Claude apps — there are no weights to download, and Anthropic hasn't signaled plans to change that. This reflects a deliberate stance on controlled deployment, not a technical gap.

xAI — Grok 2, 3, 4

Only the original Grok-1 (314B) was ever open-weighted, and it hasn't been updated since. Every Grok release after that — including the models actually in use today — remains closed and API/app-only.

Google — Gemini

Gemini itself (1.5/2.x/3) is closed. Google's open contribution is the separate Gemma line, distilled from Gemini research but not the same weights.

Amazon

No fully open-weight frontier model of its own. Amazon's role has mostly been infrastructure and investment (Bedrock hosting, Anthropic partnership) rather than releasing downloadable weights.

Running any of these offline

Weights alone aren't enough — you need an inference runtime. These four cover nearly every model on this page.

Ollama

Simplest path. One command per model (ollama run granite4), handles quantization automatically.

LM Studio

GUI app for browsing, downloading, and chatting with models locally — no terminal required.

llama.cpp

The underlying engine many tools build on. Best for custom setups, CPU-only inference, or squeezing out max performance.

vLLM

Production-grade serving for larger models across multiple GPUs — the choice once you outgrow a laptop.