The short version
Access to AI models can change without warning, so keep your own copy of a few open models on a drive you control. Pick a tier (laptop, workstation or archive drive), dry-run the download first, grab only named files from official sources, and store them with a dated record. The easiest way to do all of it is the ready-to-use agent prompt at the end of this guide.
Jump to the prompt or scan the QR code to get it right away
Hi Adopter,
Right now you can still download a capable model for free and run it on your own machine, and I wouldn’t count on that lasting. I’ll give you the exact models to download, where to get them, and how to keep them on a drive you control.
In June, a US Commerce Department order led Anthropic to turn off Fable 5 and Mythos 5 for every customer, including paying ones. Fable 5 came back on July 1. That’s one example. OpenAI discontinued the Sora app on April 26, then the Sora API on September 24. Alibaba’s Qwen3.8-Max added extra terms for certain hosted or assistant businesses, with internal use still exempt. Access can tighten because a government stepped in, or because a company changed its product or its license.
Look, the big closed labs have a reason to be uncomfortable with open weights. Their business is selling you intelligence by the token. Files you can run yourself cut against that. If they host the model, they decide who gets in. If you want intelligence independence, you have to keep a copy yourself. In July, Axios reported OpenAI and Anthropic aligning in Washington on the risks of Chinese open-weight models. Anthropic says it has never advocated for a ban. It still backed chip controls, anti-distillation measures, and mandatory testing of capable models. My view is that the posture is obvious, and those companies have more political influence than they used to. I’m not predicting a ban. I am saying if your work lives on their servers, their influence is now part of your risk.
Companies are already sending a lot of work to Chinese open-weight models. CNBC reported in September that on OpenRouter, Chinese models went from 6 to 13 percent of company tokens in February to 57 to 67 percent in the week of September 14. On Vercel, the share went from 11 percent in January to 55 percent in August. The reasons those platforms gave were price and performance, especially for coding. That’s router traffic, not proof they’re running the files in house.
Honestly, the reason I’d still keep a copy on a drive is the data. If you run the model locally, it doesn’t leave the house. Palantir CEO Alex Karp told CNBC on July 1, 2026, that technical customers want “control over their compute, their models, their data stack and their alpha.” He added, “They want to know they own the means of production. It’s not being transferred to someone else.” He was talking about control, not calling Chinese models more reliable. That’s my reading of why a local copy is worth having: you hold the file, and the data stays in house.
I’m treating this as insurance, not a forecast. Get the files while you still can. A useful start is tens of gigabytes, not a warehouse. Last checked November 14, 2026. Nvidia’s CEO transition this month is another reminder that the place most of us download from can change owners. Nvidia reports earnings on November 17, so that note is current as of writing.
Three models to start with
If you want a starting three, these are the ones I would put on a drive first. Sizes and license readings are from the Hugging Face model API, cards, and LICENSE files where present, on November 14, 2026. The weights were not downloaded, and hardware remarks are estimates.
Laptop: from unsloth/Qwen3.5-4B-GGUF, the file Qwen3.5-4B-Q4_K_M.gguf, about 2.74 GB, Apache-2.0 on the card. It is small enough to store, and likely to run, on a 16 GB-class machine.
Workstation: from unsloth/Qwen3.8-27B-GGUF, the file Qwen3.8-27B-UD-Q4_K_M.gguf, about 16.56 GB, Apache-2.0 on the card. This is the heavier general model for a 24 GB-class machine, used on its own rather than beside the large coding file.
Archive: zai-org/GLM-5.3-Flash, current tree about 328.37 GB, with an MIT LICENSE file on the repo. This is a storage-first flagship, not a file you should expect to load on a 24 GB GPU.

Sources for the opening
June shutdown: Anthropic’s Fable and Mythos access and Fable 5 redeployment notes. Sora dates: OpenAI’s help page. Labs in Washington: Axios, July 22, and Anthropic’s open-weights position. License tightening on hosted use: SCMP on Qwen3.8-Max. Adoption data: CNBC, September 26. Karp quotes: CNBC, July 1. Nvidia transition: Nvidia’s CEO announcement.
Paid subscribers get the rest of the kit: a short guide to the laptop, workstation, and archive tiers, the full model list with links, and one ready-to-use prompt for an AI agent that can dry-run a download onto a separate drive.
What the guide covers
You’ll pick a laptop, workstation, or archive kit, store and verify the files, and keep a way to run them later. I have not executed the prompt at the end, and I have not benchmarked these files.
If this is for personal use or internal work, use a commercial-safe screen. That’s a screen, not a legal sign-off. On November 14, 2026, it meant the card, and the LICENSE file where one was present, declared Apache-2.0 or MIT, the repository was not gated, and the files can be stored on ordinary disks. You still have to read each license.
Storage and runtime are different budgets. You can archive more than one computer can load at once. On a 16 GB-class or 24 GB-class machine, run one heavier model at a time. Hardware remarks below are estimates.
Pick a tier and a disk budget
Pick the laptop kit, the workstation kit, or an archive drive. Leave about 32 GB free for the laptop set, about 128 GB free for the workstation set, and a 2 TB drive if you want the large optional trees, or 4 TB if you’ll also keep some official source trees. Stop if free space minus a reserve is below that budget.
Download named files, not whole repositories
Stick to official Qwen repositories and official-family Unsloth packs, not third-party distills. Sizes use GB as 10^9 bytes, from the current Hub tree on November 14, 2026. Unless I note a LICENSE file on main, an Apache-2.0 or MIT declaration is from the card only. Use hf download with a repository-specific --include pattern, and prefer hf download --dry-run first. Those flag names were checked against official Hub docs on November 14, 2026, not a live run. Do not use huggingface-cli or --resume-download unless your installed hf download --help still lists that flag. Do not use a global *Q4_K_M* filter.
Verify the files and keep a record
Put the kit on a drive you control. Do not store access tokens in that folder, git, or the manifest. If a repository is gated, complete hf auth login and the Hub terms in your own browser. For each accepted file, keep the weights or installer, the card or README as saved that day, the LICENSE file if one exists, a note when the card declared a license with no LICENSE file on main, a sha256 if you can compute one, and the filename and byte size. After a download, check counts and bytes against the dry-run. Write a dated manifest.json and a download-log.txt with no secrets. Do not delete an older working file until the new one verifies.
Review the bag at least once a quarter, and after a release or license change that affects a file you depend on. A stale file with a checksum and a license note beats a bookmark.
Keep a way to run the files later
Store current official runtimes next to the models. llama.cpp is under the MIT License, and version 1.0.0 was published on November 13, 2026. Ollama is also MIT, as read on November 14, 2026. LM Studio is not an MIT or Apache app. The Desktop App Terms dated August 23, 2026 grant a license for personal and/or internal business purposes, subject to those terms. That’s not legal advice. Re-read the current terms before a workplace install, confirm installer sizes, and record a checksum.
The full model set
A laptop kit
This one is for a 16 GB-class machine. Required named files totaled about 12.63 GB on November 14, 2026. Optional speech generation adds 4.54 GB, for about 17.17 GB. The machine may run the 4B or 7B file, not every model at once.
General, also usable as the small and fast model: from unsloth/Qwen3.5-4B-GGUF, Qwen3.5-4B-Q4_K_M.gguf, about 2.74 GB. Apache-2.0, card only. Official Qwen/Qwen3.5-4B had an Apache-2.0 LICENSE on main and a 9.34 GB tree. Take small license or tokenizer files from the official repo if needed. It is image-text-to-text and uses the vision pair below.
Coding: from Qwen/Qwen2.5-Coder-7B-Instruct-GGUF, only the single file qwen2.5-coder-7b-instruct-q4_k_m.gguf, about 4.68 GB, lowercase. That official GGUF repository had an Apache-2.0 LICENSE file on main. Do not also pull the 00001-of-00002 shards.
Vision: from Qwen/Qwen3-VL-4B-Instruct-GGUF, Qwen3VL-4B-Instruct-Q4_K_M.gguf, about 2.50 GB, plus mmproj-Qwen3VL-4B-Instruct-Q8_0.gguf, about 0.45 GB. Vision GGUFs need the projector. Apache-2.0, card only.
Embeddings: from Qwen/Qwen3-Embedding-0.6B-GGUF, Qwen3-Embedding-0.6B-Q8_0.gguf, about 0.64 GB. Apache-2.0, card only. There’s no Q4_K_M file in this repo.
Speech recognition: openai/whisper-large-v3-turbo, current tree about 1.62 GB. MIT, card only. Plan from the current tree, then confirm with a dry-run.
Optional speech generation: Qwen/Qwen3-TTS-12Hz-1.7B-Base, about 4.54 GB. Apache-2.0, card only.
A workstation kit
This one is for a 24 GB-class machine. Required named files totaled about 47.70 GB that day. An optional music pack adds 10.09 GB, for about 57.79 GB. Laptop-only files add about 10.37 GB. Do not load the 27B general file and the 18.56 GB coding file at the same time on 24 GB.
General: from unsloth/Qwen3.8-27B-GGUF, Qwen3.8-27B-UD-Q4_K_M.gguf, about 16.56 GB. Apache-2.0, card only. Official Qwen/Qwen3.8-27B had an Apache-2.0 LICENSE and a 55.59 GB tree. Take small tokenizer or license files from the official repo if needed, not the full BF16 shards.
Coding: from unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF, Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf, about 18.56 GB. Apache-2.0, card only. Official Qwen/Qwen3-Coder-30B-A3B-Instruct had an Apache-2.0 LICENSE on main. The 18.56 GB file is tight on a 24 GB GPU with long context.
Vision: from Qwen/Qwen3-VL-8B-Instruct-GGUF, Qwen3VL-8B-Instruct-Q4_K_M.gguf, about 5.03 GB, plus mmproj-Qwen3VL-8B-Instruct-Q8_0.gguf, about 0.75 GB. Apache-2.0, card only. Vision GGUFs need the projector.
Embeddings, speech recognition, and speech generation: the same 0.64 GB, 1.62 GB, and 4.54 GB files as in the laptop kit. TTS is required here, optional on the laptop.
Optional music: ACE-Step/Ace-Step1.5, about 10.09 GB. MIT, card only. The README’s commercial-music claim was not audited.
An archive drive kit
Start with the workstation set and the laptop-only files, around 60 to 70 GB once optionals are included. Add large permissive models only after a dry-run and a disk budget. These are storage-first. Several will not run on a 24 GB GPU.
zai-org/GLM-5.3-Flash. MIT LICENSE file present. Current tree about 328.37 GB.
deepseek-ai/DeepSeek-V4.1-Flash-1017. MIT LICENSE file present. Current tree about 510.31 GB. Unsloth’s GGUF README says 128 GB or more of RAM for the smallest 1-bit. That’s published guidance, not a test I ran.
Wan-AI/Wan2.2-T2V-A14B and Wan-AI/Wan2.2-I2V-A14B. Apache-2.0, card only. Current trees about 126.20 GB and 126.21 GB.
OpenMOSS-Team/MOSS-Image-Base. Apache-2.0 LICENSE file present. Current tree 27.23 GB.
Official Qwen/Qwen3-Coder-Next-GGUF Q4_K_M only, four shards that summed to about 48.41 GB. Include every shard if you take this. Apache-2.0, card only. Do not unfilter the Unsloth pack of the same name, whose current tree was 1.80 TB and usedStorage was 6.50 TB.
Both default kits plus those large optional trees would be about 1.2 TB, not 13 TB.
What stays off the default list
If a license is restricted, keep it in a labeled folder, or off the drive, unless you’ve read and accepted the terms. That includes the Qwen Research License on Qwen/Qwen-Image-2.1, the Qwen Community License 1.0 on Qwen/Qwen3.8-Flash-Next, the Fish Audio Research License on fishaudio/s2-pro, CC-BY-NC-4.0 on Comfy-Org/YuE2, the Tencent Hunyuan Community License on tencent/HunyuanVideo-1.5, the FLUX Kommunity License on the gated Klein pack, and the LTX-2.x Community License on gated Lightricks/LTX-2.5. Several have non-commercial terms, revenue caps, or territory limits. Hunyuan’s license does not apply in the EU, the UK, or South Korea. On Qwen3.8-Flash-Next, a global *Q4_K_M* filter matches MTP helpers, not the main model.
Abliterated variants stay off the default list. If you want one, opt in after a warning.
Nemotron and some other Nvidia open models use OpenMDW or an Nvidia open-model license, not MIT or Apache, so they’re not in this default bag. If you add one later, read that card’s license.
I did not download the weights, so I did not verify transfer checksums or benchmark VRAM or RAM. Some cards declared MIT or Apache-2.0 with no LICENSE file on main. Training-data provenance was not audited. License screens are not a substitute for counsel. Cards, file names, and sizes can change after November 14, 2026. Keep your own dry-run manifest as the record.
Kit measurements and licenses are linked above and in the November 14, 2026 research notes. CLI flags were checked against the official Hub CLI and download guides. Runtime licenses: llama.cpp, the v1.0.0 announcement, Ollama, and LM Studio’s Desktop App Terms dated August 23, 2026.
The ready-to-use agent prompt
Here’s the prompt. Paste it on the machine that will hold the kit, on a separate hard drive. Tell the agent the tier you want, then confirm the path and disk budget before any weights move. It should stop after the dry-run and wait for you to approve the exact list.
Download only what I have chosen for this kit, can store, can later host from a separate hard drive, and will accept under each current license. First confirm the tier I want: laptop, workstation, or archive drive. If I have not named a tier, stop and ask. Do not download weights until a dry-run is compared with my disk budget and I approve that exact set. Do not accept gated terms, click license checkboxes, or store access tokens in the kit folder, git, chat, or the manifest. If a repo is gated, stop and ask me to complete `hf auth login` and the Hub terms in my own browser. Do not invent IDs, file names, sizes, or licenses; skip failed checks. Do not add abliterated or restricted-license models unless I opt in after a warning and will read each license. Use `hf download --dry-run` first. Do not use `huggingface-cli` or `--resume-download` unless `hf download --help` still lists that flag. Flag names were checked against official Hub docs on 14 November 2026, not a live run. Confirm the destination path on the separate drive, the disk budget, an unused reserve, and whether to store official runtime installers. Stop if free space minus the reserve is below the budget. Named files below passed a commercial-safe screen on 14 November 2026, not a legal sign-off. Download only the named files for the chosen tier, with a repository-specific `--include`. Do not use a global `*Q4_K_M*` filter. Do not silently add files from another tier.
For each repo, check the current card and license and that the include pattern matches real filenames, including every shard. If a repo is gone, renamed, gated unexpectedly, or the pattern matches nothing, stop and tell me. Dry-run the chosen tier, and if the total exceeds free space or my budget, cut optionals and dry-run again. Do not download until I approve the exact list and total, then re-run without `--dry-run`. If I asked to store runtimes, wait until after that approval. llama.cpp and Ollama were MIT as of 14 November 2026. Fetch LM Studio only after I confirm I have read the current app terms (23 August 2026 text grants personal and/or internal business use). Write a no-token log and manifest with licenses, sizes, and sha256 checksums where possible, plus the card and LICENSE file if one exists. Verify against the dry-run. Do not delete an older working file until the new one verifies. Report what was downloaded, skipped, will not run on the hardware class for that tier, and which licenses still need a human read.
Laptop tier, if that is the one I chose. Required named files, about 12.63 GB that day, for a 16 GB-class machine. Recommend 32 GB of free disk. Run one heavier model at a time.
1. `unsloth/Qwen3.5-4B-GGUF`: `Qwen3.5-4B-Q4_K_M.gguf`, 2.74 GB. Apache-2.0, card only. Take small license or tokenizer files from official `Qwen/Qwen3.5-4B` if needed. Do not pull `mmproj-F16.gguf` unless I opt in.
2. `Qwen/Qwen2.5-Coder-7B-Instruct-GGUF`: only `qwen2.5-coder-7b-instruct-q4_k_m.gguf`, 4.68 GB, lowercase. Official Apache-2.0 LICENSE on `main`. Do not also pull the `00001-of-00002` shards.
3. `Qwen/Qwen3-VL-4B-Instruct-GGUF`: `Qwen3VL-4B-Instruct-Q4_K_M.gguf`, 2.50 GB, plus `mmproj-Qwen3VL-4B-Instruct-Q8_0.gguf`, 0.45 GB. Apache-2.0, card only. Vision GGUFs need the projector.
4. `Qwen/Qwen3-Embedding-0.6B-GGUF`: `Qwen3-Embedding-0.6B-Q8_0.gguf`, 0.64 GB. Apache-2.0, card only. No `Q4_K_M` file in this repo.
5. `openai/whisper-large-v3-turbo`, current tree 1.62 GB. MIT, card only.
Optional if I opt in and disk remains: `Qwen/Qwen3-TTS-12Hz-1.7B-Base`, 4.54 GB, Apache-2.0, card only. With TTS, about 17.17 GB.
Workstation tier, if that is the one I chose. Required named files, about 47.70 GB that day, for a 24 GB-class machine. Recommend 128 GB of free disk. Do not load the 27B general file and the 18.56 GB coding file at the same time on 24 GB.
1. `unsloth/Qwen3.8-27B-GGUF`: `Qwen3.8-27B-UD-Q4_K_M.gguf`, 16.56 GB. Apache-2.0, card only. Take small license or tokenizer files from official `Qwen/Qwen3.8-27B` if needed, not the full 55.59 GB tree.
2. `unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF`: `Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf`, 18.56 GB. Apache-2.0, card only. Official `Qwen/Qwen3-Coder-30B-A3B-Instruct` had an Apache-2.0 LICENSE on `main`. Fit on 24 GB is an estimate and is tight with long context.
3. `Qwen/Qwen3-VL-8B-Instruct-GGUF`: `Qwen3VL-8B-Instruct-Q4_K_M.gguf`, 5.03 GB, plus `mmproj-Qwen3VL-8B-Instruct-Q8_0.gguf`, 0.75 GB. Apache-2.0, card only. Vision GGUFs need the projector.
4. `Qwen/Qwen3-Embedding-0.6B-GGUF`: `Qwen3-Embedding-0.6B-Q8_0.gguf`, 0.64 GB. Apache-2.0, card only.
5. `openai/whisper-large-v3-turbo`, current tree 1.62 GB. MIT, card only.
6. `Qwen/Qwen3-TTS-12Hz-1.7B-Base`, 4.54 GB. Apache-2.0, card only.
Optional if I opt in and disk remains: `ACE-Step/Ace-Step1.5`, 10.09 GB, MIT, card only. The README’s commercial-music claim was not audited. With ACE-Step, about 57.79 GB. If I also want the laptop-only files, add `Qwen3.5-4B-Q4_K_M.gguf`, `qwen2.5-coder-7b-instruct-q4_k_m.gguf`, and the VL-4B pair, about 10.37 GB more.
Archive drive tier, if that is the one I chose. This kit is storage-first. Several large optional models will not run on a 24 GB GPU. Do not unfilter `unsloth/Qwen3-Coder-Next-GGUF`, whose current tree was 1.80 TB and `usedStorage` was 6.50 TB.
Start with these default files, around 60 to 70 GB once optionals are included: `Qwen3.5-4B-Q4_K_M.gguf` (2.74 GB) from `unsloth/Qwen3.5-4B-GGUF`; only `qwen2.5-coder-7b-instruct-q4_k_m.gguf` (4.68 GB, lowercase, no shards) from `Qwen/Qwen2.5-Coder-7B-Instruct-GGUF`; the VL-4B Q4 plus Q8 mmproj pair (2.50 and 0.45 GB) from `Qwen/Qwen3-VL-4B-Instruct-GGUF`; `Qwen3.8-27B-UD-Q4_K_M.gguf` (16.56 GB) from `unsloth/Qwen3.8-27B-GGUF`; `Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf` (18.56 GB) from `unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF`; the VL-8B Q4 plus Q8 mmproj pair (5.03 and 0.75 GB) from `Qwen/Qwen3-VL-8B-Instruct-GGUF`; `Qwen3-Embedding-0.6B-Q8_0.gguf` (0.64 GB) from `Qwen/Qwen3-Embedding-0.6B-GGUF`; `openai/whisper-large-v3-turbo` (1.62 GB); and `Qwen/Qwen3-TTS-12Hz-1.7B-Base` (4.54 GB). Optional if I opt in: `ACE-Step/Ace-Step1.5`, 10.09 GB.
Add from this list only if I opt in after a dry-run and a disk budget. Do not add it silently. The default files plus all of these would be about 1.2 TB, not 13 TB.
- `zai-org/GLM-5.3-Flash`, MIT LICENSE file, 328.37 GB.
- `deepseek-ai/DeepSeek-V4.1-Flash-1017`, MIT LICENSE file, 510.31 GB. Unsloth’s GGUF README says 128 GB or more of RAM for the smallest 1-bit, published guidance, not a test I ran.
- `Wan-AI/Wan2.2-T2V-A14B` and `Wan-AI/Wan2.2-I2V-A14B`, Apache-2.0, card only, 126.20 GB and 126.21 GB.
- `OpenMOSS-Team/MOSS-Image-Base`, Apache-2.0 LICENSE file, 27.23 GB.
- Official `Qwen/Qwen3-Coder-Next-GGUF` Q4_K_M shards under `Qwen3-Coder-Next-Q4_K_M/`, 48.41 GB. Include every shard. Apache-2.0, card only.Adapt and Create, Kamil




Definitely time to save some backups of open source!