30 Powerful LLMs You Can Run for Free in 2026

the free LLM landscape in 2026, sorted by the one line that actually matters

every “free LLM” list collapses two completely different things into one word.

there’s free as in someone else runs the model and you call it: google, groq, openrouter hand you an API key at no cost. and there’s free as in the weights cost nothing and you bring the hardware: you download qwen or llama and run it on your own machine.

these aren’t variations on a theme. they’re opposites.

the hosted route costs you nothing up front but bills you in a different currency: your prompts. most free tiers train on what you send them unless they explicitly say otherwise.

the self-host route is the reverse: fully private, nothing leaves your machine, but you pay in VRAM and electricity instead of data.

so the useful question was never “what’s free.” it’s “free in which sense, and what’s the hidden cost.” sort the whole landscape by that one line and it stops being a wall of names and starts being a decision.

here’s the full map: hosted platforms you can call today, and open-weight models you can run yourself, with the catches left in instead of buried.

first, the part everyone skips: “open” doesn’t mean what you think

before the list, one distinction that trips up almost everyone: including most of the lists you’ll find ranking above this one.

open-weight means the weights are downloadable. you can run the model, fine-tune it, deploy it. that’s it. open-source means the weights and the training code and the data and the recipe are all public, you could rebuild the model from scratch.

almost everything people call “open-source AI” is actually just open-weight. of the 20 model families here, exactly one is fully open-source by that strict definition: olmo, from allen ai.

maybe two if you’re generous about granite. the rest: qwen, llama, deepseek, all of them, give you the weights and keep the kitchen door shut.

this matters because “open” hides license traps. llama is open-weight but carries a 700-million-user cap. command r is downloadable but non-commercial, free to tinker with, not free to build a business on.

gemma’s license restricts using it to train competing models. read the card, not the headline.

second: self-hosting is free, but the hardware isn’t.

the model costs nothing. the GPU does. rough rule for running a model locally at the standard 4-bit quantization: about 0.6 GB of VRAM per billion parameters.

so an 8-billion-parameter model fits a cheap 8GB card, a 32B model wants a 24GB card (a used 3090 is the value pick), and a 70B model needs two of them or a 64GB mac.

one trap inside the trap: for mixture-of-experts models, the memory tracks the total size, not the “active” parameter count the marketing leads with. a model that advertises “17B active” can still demand 55GB, because every expert has to sit in memory waiting its turn.

with those two filters: open-weight vs open-source, and what the hardware actually costs, the list reads cleanly. starting with the no-hardware route: the hosted platforms.

how to read the hardware column

all VRAM figures assume Q4_K_M — the community-standard 4-bit quantization that keeps ~95% of full quality. rule of thumb: ~0.6 GB VRAM per 1B parameters at Q4. Apple Silicon counts unified memory as VRAM (a 32GB Mac ≈ a 24GB GPU for this purpose).

MoE catch: memory tracks total params (all experts stay loaded), not the active count. a 109B MoE that activates 17B still needs ~55GB.

open-weight families — free to self-host (20)

Chinese-origin

1. Qwen (Alibaba) — Apache 2.0. the most versatile family; ships everything from 0.6B edge models to 200B+ MoE flagships. strong multilingual, toggleable thinking mode. note: the newest Qwen3.7 Plus/Max went closed/API-only — the Qwen3 / 3.5 / 3.6 lines stay open.→ practical pick: Qwen3 8B (entry, ~5.5GB) or the 27B dense / 32B (power, ~18GB on 24GB).

2. DeepSeek — MIT. reasoning-heavy; returns a chain-of-thought before its answer. the real R1 is a 671B MoE (datacenter only, ~370GB). the small “deepseek-r1:7b/14b” tags are distillations of Qwen/Llama, not the real model. → practical pick: R1 distill 14B (mid) or 32B (power); full V4/R1 is datacenter.

3. GLM / ChatGLM (Z.ai / Zhipu) — MIT. GLM-5.x leads several open coding rankings. large MoE (744B class) at the top, smaller GLM-Edge variants for consumer hardware. → practical pick: GLM-Edge (entry/mid); flagship is datacenter.

4. Kimi K2 (Moonshot) — Modified MIT. frontier coding, trillion-param MoE (~32B active). genuinely strong but needs serious hardware to self-host (~550GB+). → practical pick: datacenter / multi-GPU only. use a hosted route for casual use.

5. MiniMax M3 — open-weight. multimodal (text+image+video), 1M context, MSA architecture. coding-focused. → practical pick: datacenter-class; check Ollama for quantized community builds.

6. Yi (01.AI) — Apache 2.0. bilingual Chinese/English, 6B/9B/34B, 200K-context variants. development has slowed vs Qwen/DeepSeek — check current benchmarks before adopting. → practical pick: Yi 9B (entry) or 34B (power).

7. Baichuan — Chinese open-weight family, enterprise focus. mixed licensing — check the specific model card. → practical pick: 7B–13B class, mid tier.

8. InternLM (Shanghai AI Lab) — open releases, strong on reasoning/long-context. various sizes. → practical pick: 7B–20B class, entry/power.

9. Ernie (Baidu) — some open releases, mixed licensing. confirm the current flagship is openly downloadable before planning around it. → practical pick: verify per-model; smaller variants are entry/mid.

10. Hunyuan (Tencent) — open releases exist, mixed licensing — same caveat as Ernie.

→ practical pick: verify per-model.

Western / other

11. Llama (Meta) — open-weight but not OSI open-source: Meta community license, with a 700M-monthly-active-user cap that matters only for very large products. most downloaded family overall. Llama 4 Scout offers 10M-token context.

→ practical pick: Llama 3.x 8B (entry) up to 70B (workstation); Scout is ~55GB MoE.

12. Gemma (Google) — runs well on modest hardware; Gemma 3/4 add vision and tool calling. Gemma 4’s 12B fits in 16GB; the 26B MoE hits ~85 tok/s on consumer hardware. license restricts fine-tuning for competing models,read it.

→ practical pick: Gemma 4 12B (mid) or 26B (power).

13. gpt-oss (OpenAI) — Apache 2.0. OpenAI’s open-weight family; not served through the OpenAI API, you download and run it. gpt-oss 20B is the “16GB sweet spot”; the 120B needs ~60–65GB.

→ practical pick: gpt-oss 20B (mid/power), 120B (datacenter).

14. Mistral / Devstral — Large 3 and Small 4 now ship Apache 2.0 (a shift from earlier restrictive licensing). Small 4 packs Devstral’s agentic coding into a ~6B-active package; Mistral Small 24B owns the function-calling/JSON niche.

→ practical pick: Mistral Small (entry/mid), Devstral for agentic coding.

15. Phi (Microsoft) — MIT. small, punchy reasoning models (~1.5B–14B); “quality data over quantity.” Phi-4-mini runs on a mini PC without a discrete GPU.

→ practical pick: Phi-4 / Phi-4-mini, entry tier.

16. Nemotron (NVIDIA) — open-weight, efficient inference; Nemotron 3 line. hybrid architectures (Mamba layers).

→ practical pick: varies by size; mid to datacenter.

17. OLMo (Allen AI) — Apache 2.0, and one of only two truly open-source families (weights + training code + data + checkpoints, fully reproducible). research-grade; competitive at size but trails Qwen/DeepSeek on leaderboards. largest is ~32B.

→ practical pick: OLMo 2 7B/13B, entry/mid.

18. Falcon (TII, UAE) — Falcon license (based on Apache 2.0); free under $1M revenue, 10% royalty above. Falcon-H1 uses a hybrid SSM+attention design, 256K context across sizes, 1B–34B.

→ practical pick: Falcon-H1 7B–34B, entry to power.

19. Granite (IBM) — Apache 2.0, enterprise/RAG-focused. small long-context MoE variants (1B–3B) for low latency, plus 8B–70B. runs on Apple Silicon down to 16GB.

→ practical pick: Granite 8B (entry), bigger for enterprise.

20. Command R (Cohere) — open weights, **but non-commercial license: **free to use and experiment, not free for your business. enterprise RAG strength. (Tiny Aya 3.35B is also CC-BY-NC, 70+ languages.)

→ practical pick: fine for personal/research; needs a commercial license otherwise.

free hosted platforms — no hardware needed (10)

these give you an API key (or chat UI) at no cost, capped by rate limits. the canonical living list of current quotas is the GitHub repo cheahjs/free-llm-api-resources, limits shift weekly, so verify before building.

catch: most train on your data unless stated otherwise; only self-hosting is fully private.

21. Google AI Studio (Gemini) — best free access to a frontier closed model. ~1,500 requests/day on Gemini Flash, no credit card, no expiry (resets daily, not a trial). 1M context, handles images/PDFs. free-tier prompts may train Google’s models, keep sensitive data off.

22. Groq — fastest free option; runs open-weight models (Llama, Qwen, Kimi, gpt-oss) on LPU hardware at 300+ tok/s. concrete caps, e.g. ~30 req/min, 1,000/day on a 70B model. clearer no-training policy.

23. Cerebras — like Groq, very fast open-weight inference on wafer-scale chips; generous free tier, no-training policy.

24. OpenRouter — widest variety through one key; 25+ permanently free models (filter with the :free suffix), no credit card, failover routing. clear no-training option.

25. GitHub Models — free for development inside rate limits; mixed catalog (OpenAI, Llama, Mistral, DeepSeek). good if you live in the GitHub/Copilot workflow.

26. Cloudflare Workers AI — edge inference, 10,000 neurons/day free. good for serverless apps; overages are cheap but not unlimited.

27. Mistral (La Plateforme) — free developer/experiment tier. catch: the Experiment tier requires opting into training to access the 1B-tokens/month quota.

28. Hugging Face Inference — thousands of models; serverless inference limited to models under ~10GB, with tight rate limits and cold-start latency. best for trying unusual or brand-new models.

29. NVIDIA NIM — hosted open models with a free tier, but generally requires billing setup and leans toward trial-style credits. treat as trial, not permanent-free.

30. Together AI — free models plus ~$1–25 signup credits. the credits are a trial, not a permanent free tier, budget accordingly.

~m0h