The Best Downloadable AI Models of 2026, Matched to the Machine You Already Own

Jensen Huang resigned from Nvidia on November 12, and nobody knows what happens to the company’s nearly $13 billion bet on open-weight AI. Below: the models worth installing this month, sorted by the memory you already have, and what the Hugging Face uncertainty actually means for you.

This piece is the second entry in our local-AI series; if you want the hardware fit and cost math first, start with Part 1: who should actually run a local LLM in 2026.

What Just Happened at Nvidia

On November 12, 2026, Jensen Huang resigned as CEO of Nvidia. CFO Colette Kress is interim CEO. Per TechCrunch’s reporting, the resignation followed Huang’s admission of an affair with an alleged Chinese spy, which has reframed his open-weight advocacy as a national-security question in Washington.

That matters because Huang was the internal champion of the two biggest open-weight commitments Nvidia has made. The first is the Nemotron family of open models. The second is the $12.93 billion acquisition of Hugging Face, announced September 3 and scheduled to close in the first half of 2027, which would be Nvidia’s second-largest acquisition after its $20 billion purchase of Groq’s assets in December 2025. One Nvidia insider told TechCrunch they don’t expect Nemotron development to change. That is a single anonymous source, so treat it as one data point, not a signal.

Huang’s departure also puts a political artifact in a new light. On July 24, Nvidia published the “Open Weights and American AI Leadership” letter, signed at launch by Microsoft, Meta, Hugging Face, IBM, and Palantir and reported to have grown to roughly 50 companies within a day. The letter was a response to the policy fight ignited by Moonshot AI’s Kimi K3 release on July 16, which pushed open-weight restrictions into live Washington debate.

The counterweight so far is verbal. Hugging Face CEO Clément Delangue vowed publicly on November 12 to “keep fighting for open-source AI, harder than ever.” The Department of Justice opened inquiries into the deal’s competitive impact in October, per Reuters. Whether the deal closes, renegotiates, or unwinds is genuinely open. What is not open is the part you control, which is the rest of this article.

The Picks, by the Machine You Have

Macro close-up of a RAM memory module held above a workbench
Photo: Harrison Broadbent / Unsplash

One convention before the tables. A comfortable 4-bit run takes roughly 0.55GB of memory per billion parameters, plus 10 to 25 percent overhead for context and the runtime itself. Apple unified memory and discrete GPU VRAM count as one budget here. Every minimum below includes modest context headroom, and every figure was checked against that arithmetic.

Table figures are as of November 13, 2026, and every link in this piece was checked November 2026.

8GB: The Everything-You-Need Tier

A Raspberry Pi 4 single-board computer on a workbench, viewed from above
Photo: Vishnu Mohanan / Unsplash

If your machine has 8GB, one install covers almost everything: Qwen3.5-9B, the best measured performer at this tier in SOTAAZ’s coding tests, with a 262K context window and a vision encoder built into the base model, so it doubles as the tier’s document-and-image reader. It is Apache 2.0.

A quick note on the last column before the tables. Suggested context is the conversation length that actually fits in the memory left over after the operating system, your browser, and your other apps take their share. On an everyday machine that leaves roughly 3GB for context on an 8GB computer, 6GB on 16GB, 14GB on 32GB, and 24GB or 40GB on a 96GB or 128GB workstation. Those figures are the pool before the model’s own weights, and the weights come out of the same memory: a 6GB model on an 8GB machine has far less room for context than a 4GB model does, and the table rows are computed exactly that way.

The low end of each range assumes normal multitasking; the high end assumes a lean machine with little else running. Figures assume the default 16-bit KV cache, are capped at each model’s native context window, and are cross-checked against the model weights so tier memory is never exceeded. Switching to an 8-bit KV cache (q8, supported by llama.cpp, Ollama, and LM Studio) roughly doubles the high end. One corollary worth stating plainly: the very long context windows quoted for cloud models, 200,000 tokens and more, are exactly what does not fit at the small tiers.

ModelUse caseParams / quant sizeMin memorySuggested context
Qwen3.5-9BGeneral chat, light coding, vision9B / ~6GB Q4_K_M8GB7k–16k
Qwen3-VL 8BDedicated OCR and document work8B / ~6.2GB Q4_K_M8–12GB1k–2k (tight)
Gemma 4 E4BOn-device and edge assistant8B total, 4B effective / 4.5GB Q46–8GB66k–128k
Qwen3.5-4BUltra-light agents, function calling4B / ~3.3GB Q46GB44k–94k

Two notes. Qwen3-VL 8B is the stronger pick if your workload is scanned documents, receipts, and charts, with dynamic resolution up to 4096 pixels. Gemma 4 E4B runs on a Raspberry Pi 5 via LiteRT (measured 1,432MB peak memory, 9 tokens per second), which makes it the tier’s phone-and-Pi option.

16GB: Reasoning and Your First Coding Agent

The 16GB tier adds real reasoning and one of the most interesting releases of the year: gpt-oss-20b, the first open-weight model OpenAI has shipped since GPT-2. It runs in 16GB and is Apache 2.0. One caveat, straight from OpenAI’s own documentation: gpt-oss models use the “harmony” response format, and they misbehave in harnesses that don’t implement it. If you run them through LM Studio or the OpenAI-compatible path in Ollama, you get the format handled for you.

ModelUse caseParams / quant sizeMin memorySuggested context
gpt-oss-20bGeneral assistant, reasoning21B MoE, 3.6B active / 12.8GB MXFP416GB32k–70k
Gemma 4 12BCreative writing, chat12B / 6.7GB Q4 (official)12–16GB40k–88k
Devstral Small 2 24BCoding, agent loops24B / ~13.5GB Q416GB (tight)~2k–6k (tight)
Z-Image-TurboImage generation6B / ~7GB FP816GBn/a — image model

Gemma 4 12B has the highest prose quality per gigabyte under 24GB, per Google’s official memory table. Devstral Small 2, Mistral’s coding line, fits 16GB with an 8-bit KV cache and is the cheapest entry into agentic coding loops. For images, Z-Image-Turbo (6B, Apache 2.0) generates in about a second on a 16GB GPU and is, per the October 2026 rankings, the best local image model for most people. FLUX.2 [klein] 4B (~13GB) is the other permissive option; FLUX.1 dev needs 24GB at full quality and carries a non-commercial license.

32GB: The Sweet Spot

This is the tier where local models stop feeling local. Qwen3.8-27B, released August 2026, is the current all-rounder: 256K context and a 91.1 OmniDocBench score for document work. The Gemma 4 pair covers speed and prose.

ModelUse caseParams / quant sizeMin memorySuggested context
Qwen3.8-27BGeneral all-rounder, OCR27B / ~18GB Q424–32GB92k–192k
Gemma 4 26B-A4BFast assistant (MoE)26B, 3.8B active / 14.4GB Q4 (official)24GB160k–256k
Gemma 4 31BCreative writing flagship31B / 17.5GB Q4 (official)24GB36k–76k
Qwen3.5-27BCoding (non-agentic)27B / ~17GB Q424GB98k–208k
Nanbeige4.3-33BAgentic coding33B / 20.8GB Q4_K_M32GB22k–46k

Gemma 4 26B-A4B is a mixture-of-experts model that runs exceptionally fast and debuted at number six among open models on Arena text. Gemma 4 31B tops the independent creative-writing rankings for this tier (EQ-Bench Creative Writing v3 around 16.0, effectively tied with Qwen3.8-27B). Nanbeige4.3-33B earns its 32GB requirement: it holds the best independent Terminal-Bench 4.0 score (29.8%) of anything that fits a single 32GB card.

64GB and Up: The Mac Studio Tier

Above 64GB, you can run the largest fully permissive open models in existence today. Mistral Small 4 (119B MoE, 256K context, vision and reasoning built in) is the flagship, and it needs a 96GB-class machine at 4-bit. Some secondary guides claim it fits in 16–32GB; that is arithmetically impossible at this quant, and you should trust the arithmetic, not the guide.

ModelUse caseParams / quant sizeMin memorySuggested context
Mistral Small 4Frontier-class local assistant119B MoE / ~65GB Q496GB18k–42k
FlintLocal coding agent39.8B MoE, 2.5B active / 26GB INT448GB256k (native cap)
gpt-oss-120bReasoning heavyweight117B MoE, 5.1B active / 60.8GB MXFP480GB128k (native cap)
Qwen3.5-122B-A10BMultimodal flagship122B MoE, 10B active / ~68GB Q496GB256k (native cap)
Llama 4 Scout (flagged)Large-context multimodal109B MoE, 17B active / ~60GB Q464–96GB58k–124k

Flint, released October 21 by Recursive, is purpose-built for agentic coding and runs on Ollama, LM Studio, and llama.cpp. Its headline benchmark numbers (an 89.7% on SWE-bench Pro v2) are vendor-reported; third-party confirmations are still thin. gpt-oss-120b lands near o4-mini-class reasoning inside one 80GB envelope, which means a 128GB Mac or a single H100. If a maxed-out Mac mini doing quiet local work is the life you want, our Project N.O.M.A.D. build is the sibling worth reading after this one. Llama 4 Scout is capable but ships under the restricted Llama 4 Community License, so it earns a flag, not a recommendation; the Apache picks above it are better for almost everyone.

The Frontier You Can’t Buy

For completeness: Kimi K3 (2.8T parameters, 104B active, with a smallest usable build around 610GB even at 1-bit), GLM-5.3, DeepSeek V4.1, and Meta’s Muse Spark Open (816B) are all downloadable, and none of them fit home hardware. They want 200GB to 1.6TB of memory. The frontier exists, but it rents by the hour. Everything in the tables above is yours permanently, on your desk, for the cost of the electricity. If you want the full picture of what running without that rental gets you, we covered it in our look at AI beyond the cloud.

Licenses in One Minute

Almost everything in those tables is Apache 2.0 or MIT, which means: run it anywhere, modify it, ship products on it, no royalties. There are two traps worth knowing before you download.

Llama 4 Scout runs under the Llama 4 Community License, not an open-source license. It requires “Built with Llama” attribution, a name-prefix rule for derivatives, and it revokes rights automatically once a product built on it passes 700 million monthly active users. For a household, the attribution clause is a shrug. For a startup, the carve-out is a term sheet. Kimi K3 uses a bespoke license with a gate on hosted commercial services at $20M in revenue. Neither is a villain; both are terms you should read before you build on them, and “community license” is not a synonym for open source.

What Happens to Hugging Face Now

A rack of server equipment in a dark room, shallow depth of field
Photo: Tyler / Unsplash

Three plausible paths, and none of them changes what runs on your machine.

The Deal’s Three Paths

The acquisition could close as planned in the first half of 2027 under Kress, who inherits it as interim CEO. The terms could be renegotiated under pressure, with the Department of Justice’s October inquiry supplying the leverage; Reuters’ deal reporting put the original terms at $12.93 billion, with $11.9 billion to investors and up to $1 billion in retention. Or it could unwind entirely if the inquiries harden into action. Each path is live. What reporting does not support is assigning odds, so this piece won’t.

Your Weights Are Safe

Here is the structural fact underneath the drama: model weights are files. Hugging Face hosts 3 million of them for 18 million developers, but hosting is redundant by design now. ModelScope mirrors roughly 170,000 repositories and is the canonical home for Qwen. Ollama’s library, LM Studio’s catalog, Kaggle, GitHub releases, and the Unsloth community’s GGUF mirrors all carry major releases within days of publication. Huang himself pledged that “Nvidia compute will not be required to build on or deploy through Hugging Face,” and Nvidia has released more than 500 models there.

If you download a model today, you own the file tomorrow no matter which corporate story resolves. A company can buy a host. It cannot buy a file that is already on your disk, mirrored across five other platforms, and forkable by anyone.

Who Funds Open Weights if Nvidia Steps Back

Nvidia matters to this ecosystem, but it is one funder among many, and the list predates the acquisition. Mistral funds Mistral Small and Devstral. Alibaba funds Qwen. Google funds Gemma and published the official memory table this article’s tables lean on. OpenAI now funds gpt-oss. Meta funds Llama and Muse Spark. The Allen Institute funds OLMo end to end, weights to training data. A Swiss consortium funds Apertus. European public money funds parts of the stack outright. The July letter’s signatories did not sign a Jensen Huang personal project; they signed a policy position, and it outlives any one executive.

What to Actually Do This Week

Three things, in order:

1. Download your picks now, through Ollama or LM Studio, straight to disk. If a model on the list above interests you, the ten minutes it takes is the entire exposure you have to the deal.

2. Watch the first-half-2027 close date and the DOJ inquiry outcome if you follow the business story. The lede of this article may need updating; the tables will not.

3. Change nothing about your installed models. Nothing about them changes either way.

Wrap-Up: Why This Matters

A quiet writing desk with a notebook, pen, and laptop
Photo: Glenn Carstens-Peters / Unsplash

Every dependency in your digital life has a landlord: the password manager with a renewal date, the cloud drive with a terms-of-service update, the smart-home assistant that goes dumb when a server does. Open-weight models are the exception. The file is on your disk, the license lets you keep it, and the worst-case corporate scenario still leaves you running the exact same software you ran yesterday.

The churn at Nvidia is real and worth watching. It decides who hosts the files, not what you can run. If you have been waiting for a reason to start, the machine you own and an evening are the only requirements.

Sources

1. TechCrunch: Jensen Huang’s exit puts Nvidia’s support for open source in question (Nov 12, 2026)

2. TechCrunch: Hugging Face CEO vows to keep fighting for open-source AI (Nov 12, 2026)

3. TechCrunch: Nvidia confirms it will buy Hugging Face for $12.9 billion (Sep 3, 2026)

4. Reuters: Nvidia to buy Hugging Face in nearly $13 billion bet

5. Nvidia: Open Weights and American AI Leadership (letter PDF) (Jul 24, 2026)

6. CNBC: Nvidia agrees to buy Hugging Face (Sep 3, 2026; Groq $20B context)

7. Nvidia blog: Nvidia to acquire Hugging Face (Huang quote, 500+ models)

8. Google: Gemma 4 announcement (Apr 2, 2026)

9. Gemma 4 official memory table

10. Qwen3.5 repository (release chronology, Apache 2.0)

11. Qwen3.5-9B model card

12. OpenAI: Introducing gpt-oss (403 bot-wall; confirmed via search index)

13. gpt-oss repository (Apache 2.0, harmony format caveat)

14. Mistral Small 4 model card (119B MoE, Apache 2.0)

15. Kimi K3 model card (2.8T, Kimi K3 License)

16. Allen Institute: OLMo 3

17. Raspberry Pi: Edge AI with LiteRT and Gemma (measured Pi 5 figures)

18. Llama 4 Community License text

19. GPUwerk: Best open-weight coding models (Nov 2026 rankings)

20. Prompt Quorum: Best local LLMs for creative writing (EQ-Bench figures)

21. VDF: Best local image generation models (Z-Image-Turbo pick, FLUX.2 [klein])

22. Presenc: ModelScope vs Hugging Face 2026 (~170K ModelScope repos)

23. VentureBeat: Kimi K3’s full weights, open with a caveat (429 rate-limit; confirmed via search index)

24. Digital Applied: Local image generation with FLUX and Stable Diffusion

25. SOTAAZ: Which local coding model should you run on your GPU? (8, 12, 16 and 24 GB on 4 Python tasks)

Leave a Comment

Required fields are marked.