Meet POCKET-27B. Qwen3.8-27B in ARCHsloth Q6_K GGUF. Choose files & download · Quality & limits

🏆 SuperCLUE November 2026: #1 open-weight model in math and science reasoning

From the VIDRAFT Darwin family. Our flagship reasoning model, Darwin-180B-RSI-R3, was scored by SuperCLUE, China's most widely cited LLM benchmark, on its private November 2026 test set: 73.68 in mathematical and scientific reasoning, #1 among open-weight models and #2 overall among 19 models, behind only OpenAI's GPT-6.1-Sol and ahead of Gemini-4-Argon, DeepSeek-V4.1-Flash, Grok-4.7 and Qwen3.8-Max. The same recursive self-improvement line and the same team built the POCKET on-device builds.

Darwin-180B-RSI inherited its expert layers from Qwen3.8-Flash-Next and rebuilt only the attention and shared-expert layers through recursive self-improvement. On the same board the parent, Qwen3.8-Flash, placed 15th overall (39.81) with 57.89 in math and science reasoning; Darwin placed 9th (43.08) with 73.68.

Overall intelligence index, November 2026 (screenshot)

Source: SuperCLUE general leaderboard, November 2026 (screenshots taken November 5, 2026). Scores are SuperCLUE's, not VIDRAFT's.

POCKET-26B-GGUF

Korean-tuned Gemma4 GGUF for local text assistance, with 11 GB and 17 GB build options.

Download files · Setup guide · Benchmarks

POCKET-35B · POCKET-26B · POCKET-Darwin-180B

At a glance

This package keeps the source model unpruned and applies VIDRAFT's Korean-tuned quantization. It is not a promise of universal app support, a measured 12 GB-phone fit, or an MLX package.

Choose your build

File sizes are rounded decimal GB, not process-memory measurements. An 11 GB file does not establish reliable operation on a 12 GB phone: the operating system, context, runtime buffers and other applications also need memory. For smaller phone-oriented artifacts, see the separate POCKET-KR-GGUF card and its limits.

Download and run

Install the Hugging Face CLI and a llama.cpp build that supports Gemma4. The example downloads only the recommended file from the audited revision; it does not clone or download the whole repository.

1. Download the recommended build.

hf download FINAL-Bench/POCKET-26B-GGUF POCKET-26B-Q2_K.gguf --revision 9f49f5c276efafe909f6b2b2fb161181ce417a45 --local-dir ./POCKET-26B

2. Run on the CPU.

llama-cli -m ./POCKET-26B/POCKET-26B-Q2_K.gguf --device none -ngl 0 -t 8 -c 4096 -n 256 -p "Explain why local AI can be useful."

The expected result is generated text in the terminal, not a fixed benchmark answer. These command examples were syntax-reviewed but not executed during this review.

Commands use the standalone llama.cpp executable names. Some distributions expose a unified llama launcher; consult that distribution's help and the official model-loading documentation. Use a physical-core-based thread count appropriate to your device; a higher number is not automatically faster.

For installation, memory errors and app-specific help, use the model-specific setup guide. Model files remain directly available from Hugging Face without a VIDRAFT login.

Runtime compatibility

The same file format does not guarantee support in every app or version. A dated, exact-version compatibility matrix remains a community contribution opportunity.

Hardware and performance

These are the limits of the supplied device evidence, not estimated performance promises. The CPU-only quickstart explicitly disables GPU acceleration with --device none; -ngl 0 alone is not the strongest no-GPU guarantee in every backend.

A useful community result identifies the exact file, runtime version, device, context size, peak memory and prompt/generation rates. Do not transfer POCKET-35B's Xeon or Mac speeds to this different architecture.

Quality and evaluation

Publisher-reported GPQA Diamond results: a single greedy pass over 198 questions using the publisher's harness.

This small comparison does not establish general statistical equivalence or universally lossless quantization. Paired answer-level results, a prespecified equivalence margin, full harness details and broader tasks would be needed for a stronger equivalence claim. No dedicated Korean task-quality score is supplied in the source card; the GPQA table should not be presented as a Korean-language quality benchmark.

Quantization and unpruned-model notes

The original card describes Korean-tuned mixed precision: more sensitive weights remain at higher precision. It reports that the model remains unpruned, retaining all 128 experts, and that pruning harmed its low-bit robustness.

The original card also mentions approximately 44% for a plain uniform-Q2 comparison. The exact comparator file and full evaluation configuration were not supplied, so this is a publisher-reported observation, not an independently verified ranking or an equivalent-format guarantee.

Intended uses

• Local text assistance and Korean-language workflow experimentation on suitably provisioned devices.
• Comparing the 11 GB and 17 GB quantization options with a task-specific evaluation.
• CPU-only workflows where local execution is preferred; no cloud service is required by the supplied command.

Limitations

• Exact peak memory, reliable phone fit and on-device iPhone/Mac throughput were not supplied.
• “GGUF” does not mean every app or version supports Gemma4.
• The GPQA comparison is one 198-item greedy evaluation, not general quality equivalence.
• Korean-tuned quantization is the package's positioning; a dedicated Korean task benchmark is not supplied here.
• File size is not a complete RAM budget; longer context increases memory requirements.

License and attribution

This model card declares Apache-2.0, consistent with the current google/gemma-4-26B-A4B-it metadata and Google's official Gemma 4 license page.

Starting checkpoint: Gemma4-26B-A4B-it by Google. Korean-tuned quantization and this package: VIDRAFT / FINAL-Bench. Runtime tooling: the llama.cpp contributors. Read the license and retain required notices and attribution when redistributing.

Support and updates

For device selection and installation help, visit the VIDRAFT POCKET setup guide. Report reproducible issues in the model discussions, including the filename, runtime version, device, memory and command; do not post private prompts or credentials.

A browser-accessible POCKET-26B CPU demo is a separate hosted service; its availability and latency are not local-device measurements. No local-inference telemetry or website gate is required to obtain these public files.

POCKET family

Explore the POCKET collection. For phone-oriented derivatives, see KR GGUF, KR MLX and EN GGUF.

More POCKET projects and research

Qwen3.8 Flash-Next GGUF is a separate parent-model package, distinct from the POCKET-Darwin-180B RSI-R3 weights. Image projects are also separate: POCKET-Image-Zimage, POCKET-Zimage-CPU and Image Studio.

Research collections: Darwin, Aether, VKAE.

References

• Source card and publisher measurements, revision 9f49f5c.
• Google's Gemma4 starting checkpoint and Apache-2.0 license.
• Hugging Face CLI download documentation.
• llama.cpp build and CPU-backend guidance.
• llama.cpp model loading.
• Hugging Face with Ollama and LM Studio.
• VIDRAFT: on-device LLMs without a GPU and model quantization.

Documentation reviewed on 2026-11-04.