Local AI, with a clear starting point

Your model.
Your machine.

Meet the POCKET family: 26B, 35B and Darwin-180B GGUF models. Choose a build that matches your hardware, download directly, and run text inference locally.

Find your model Open the quickstart

No VIDRAFT account. No download gate. No inference telemetry from this guide.

01 / Choose honestlyFile size is not runtime memory.
02 / Keep access openDirect Hugging Face downloads.
03 / Stay in controlRun locally; ask for help optionally.
One family. Different requirements.

Start with your hardware.

Compare files, runtime support and available memory—not just the parameter count. Each model has its own evidence and license.

Darwin lineage · Qwen-family MoE

POCKET-35B

A choice of four quantizations for local text generation on a CPU desktop or server.

8.24–21.17 GB
GB · 4 file choices
What it is
35B-class sparse MoE. Darwin-36B-Opus lineage; qwen35moe runtime support required.
Before you download
Q2_K card guidance: 16–24 GB RAM. Q4_K_M: 32 GB. Leave extra OS and context headroom.
Evidence boundary
Publisher-reported 27.0 tok/s on Xeon, 16 threads, IQ1_M. Not a speed promise for other files or devices.
Use this model Browse model files
Model card & benchmarksTry hosted CPU demo
Direct file downloads
Korean-tuned · Gemma-family MoE

POCKET-26B

Two unpruned, Korean-tuned GGUF builds for local conversation and task evaluation.

11.12–16.80 GB
GB · 2 file choices
What it is
Gemma 4 26B-A4B lineage, about 4B active parameters. Verify Gemma 4 support in your runtime.
Before you download
File size is not RAM usage. Check free memory, app limits and context size; no universal phone-fit claim.
Evidence boundary
Publisher GPQA Diamond: Q4_K_M 67.7%; Q2_K 67.2% (198 items, greedy). No device speed verified here.
Use this model Browse model files
Model card & benchmarksTry hosted CPU demo
Direct file downloads
Darwin RSI · Qwen-family MoE

POCKET-Darwin-180B

A larger text-only model for high-memory systems or a documented SSD-streaming setup.

111.33 GB
GB total · all 4 shards required
What it is
Darwin-180B-RSI-R3 lineage. qwen4exp support; publisher requires llama.cpp b11048 or newer.
Before you download
120 GB free NVMe recommended by the card. CPU server and GPU-assisted SSD routes have different requirements.
Evidence boundary
Publisher CPU report: 18.4–21.0 tok/s on EPYC 9365, 16 threads. Laptop result uses an 8 GB GPU—not CPU-only.
License
Qwen Community License 1.0
Additional commercial conditions apply; review the full license.
Use this model Browse model files
Model card & benchmarks
Download all four shards
A practical memory check. Leave headroom for your operating system, context/KV cache and other applications. Larger context uses more memory. The Darwin-180B laptop setup streams from an NVMe SSD and uses a GPU; it is not the same as the CPU-only route. Decimal GB are download sizes, not peak RAM.
Choose → download → run

POCKET-35B · local quickstart

Read the requirements, choose your files, then copy the commands into your terminal. Nothing runs or downloads automatically.

Runtime prerequisite

A current llama.cpp build with qwen35moe support. Check your runtime version before downloading.

Official llama.cpp setup · Official releases

Memory & disk check

The model card suggests 16–24 GB RAM for Q2_K and 32 GB for Q4_K_M. Leave room for the OS and context; these are not guaranteed-fit specifications.

01 Prepare the download tool

python -m pip install --upgrade huggingface_hub
hf --help

Requires Python and pip, or install the official standalone hf CLI. Public files can be downloaded without a VIDRAFT account; this guide does not require a Hugging Face login. Follow the upstream tool's terms.

02 Download the pinned model files

hf download FINAL-Bench/POCKET-35B-GGUF "POCKET-35B-Q2_K.gguf" --revision 3870039a87406fb90764b27e3984c8b1a36b269e --local-dir "./models/pocket_35b"

Commands use a fixed public revision for reproducibility. Review the latest model card before deployment. Check free disk space first; the download tool may need temporary/cache headroom.

03 Run on your machine

llama-cli -m "./models/pocket_35b/POCKET-35B-Q2_K.gguf" --device none -ngl 0 -t 8 -c 2048 -n 256 -p "Explain three practical uses for a local language model."

CPU-only example. The --device none flag explicitly disables GPU devices; -ngl 0 sets GPU layer offload to zero. Adjust thread count to your physical cores and available memory. A short context keeps the first test modest.

Download selected GGUF directlyAsk the POCKET-35B community

Measurements, not universal promises

Know what the numbers mean.

Published model-card results, reviewed on November 4, 2026. They were not rerun for this page. Different models, devices and evaluation settings are not directly comparable.

POCKET-35B

27.0 tok/sCPU generation · Xeon · 16 threads · IQ1_M

The publisher also reports HellaSwag 61.0% on 400 questions for IQ1_M; the comparison's confidence intervals overlap. Q4_K_M and Q2_K have separate GPQA results. Neither the speed nor quality figure applies to every quantization.

Read the pinned benchmarks ↗

POCKET-26B

67.7% / 67.2%GPQA Diamond · Q4_K_M / Q2_K · 198 items, greedy

These are the publisher's own harness results. A small benchmark cannot establish universal quality parity. No on-device phone throughput is verified by this guide; app compatibility and memory limits need checking.

Read the pinned benchmarks ↗

POCKET-Darwin-180B

87.65%MMLU-Pro · 2,000 paired questions · this GGUF

The publisher reports the same sample accuracy as R3 BF16, with paired difference +0.00 points and 95% CI [−0.95, +1.00]. This is not proof of identical answers or parity on all tasks. The DGX Spark 39.25 tok/s figure in the source is for a reference build, not this exact build.

Read the pinned benchmarks ↗
Download volume is not user count. The three repositories reported 1,174,638 combined rolling-30-day downloads in the November 4, 2026 Hugging Face API snapshot—not unique people, installations, website visits or a conversion rate. Split GGUF shard requests can each count; Darwin-180B requires four shards. See Hugging Face download-count rules and the POCKET collection.
For builders and local users

Get unstuck.
No company required.

Use the relevant model's public community discussions for setup questions and reproducible bug reports. Reading is public; posting may require a Hugging Face account.

Share the runtime version, model filename, OS and a redacted error—not private prompts, documents or access credentials.

Optional · teams and partners

Bring POCKET
into your workflow.

Talk to VIDRAFT about integration, on-premises deployment, hardware collaboration or a POCKET Box project. Community help and model access do not depend on a business inquiry.

Email VIDRAFT ↗

Opens your email application; nothing is submitted by this page.

What can I use these models for?

Explore local conversation, drafting, summarization, coding assistance and research prototypes, subject to each model's license. Evaluate the selected model and quantization on your actual tasks before relying on outputs. The 26B model is Korean-tuned; Darwin-180B is a text-only GGUF.

Will a model fit my laptop or phone?

A download size does not include all runtime memory, the OS or context cache. Memory bandwidth, SSD speed, thermal limits, backend support and quantization all affect results. The three repositories here are not universal phone builds. Check the specific card and start with a small context; do not treat one measured device as certification for all devices.

Do all three models have the same license?

No. The 35B and 26B repositories declare Apache-2.0. Darwin-180B uses Qwen Community License 1.0, which includes additional commercial conditions, including a separate-license requirement for certain commercial Model-as-a-Service and AI Work Assistant businesses. Read the linked full license before deployment; this page is not legal advice.

Does this guide track my local inference?

No. This page contains no analytics SDK, collection endpoint, browser identifier or persistent storage. Its optional integration interface only emits coarse local browser events; it sends no network analytics. Copying a command does not report that it was run. Models and tools have their own code and terms; review them independently. The hosted demos are separate services, not private local inference.

What if the runtime cannot load the file?

Check the architecture and your exact llama.cpp version first: qwen35moe for POCKET-35B, Gemma 4 support for POCKET-26B, and qwen4exp with the card's b11048-or-newer requirement for Darwin-180B. Check that downloads completed and all four Darwin shards are together. Use the model's discussion page if the issue remains; do not post secrets.