L0xRE-27b-Low

A 27-billion-parameter text-generation model designed for a smaller GPU-memory footprint. The model file is 8.62 GB. The recommended starting point is an NVIDIA RTX 30, 40, or 50 series GPU with at least 12 GB of VRAM, using the L0xRE BeeLLama runtime.

“27b” is the parameter count, not the download size. “Low” describes the memory-focused model build. Context length and available GPU memory depend on the card, driver, runtime profile, and other applications.

Start here: runtime downloads, installation, and GPU test results. There is one Linux/WSL runtime archive and one Windows runtime ZIP for the supported GPU families.

Download these two files

For the recommended setup, download both files: 9.76 GB in total, plus the runtime package. The helper is called a drafter; it is an additional file, not another name or edition of L0xRE-27b-Low. The main model can also run without a drafter, with speculative decoding disabled.

The smaller Qwen3.8-27B-DFlash2-Q2_K.gguf file is an optional 705 MB drafter for memory-constrained experiments. Start with the recommended Q4_K_M drafter to match the published default configuration.

Install and run

1. Download and extract the runtime for your OS from the universal release.
2. Download the two recommended files above into the extracted runtime's models folder.
3. Verify the runtime and model checksums using the installation guide.
4. Start the server from the extracted runtime folder:

# Linux / WSL
./l0xre serve --profile 12gb -m models/L0xRE-27b-Low.gguf \
  -md models/Qwen3.8-27B-DFlash2-Q4_K_M.gguf --host 127.0.0.1 --port 8080

# Windows
.\l0xre.cmd serve --profile 12gb -m models\L0xRE-27b-Low.gguf -md models\Qwen3.8-27B-DFlash2-Q4_K_M.gguf --host 127.0.0.1 --port 8080

When ready, the server's health endpoint is http://127.0.0.1:8080/health and its API base is http://127.0.0.1:8080/v1. For a compatible chat client, use that API base and select L0xRE-27b-Low from the server's /v1/models endpoint.

This file requires the L0xRE runtime. Direct use with stock llama.cpp, Ollama, LM Studio, Transformers, or other inference engines is outside this release's supported setup. The GGUF extension alone does not establish compatibility.

GPU support and context

The runtime targets NVIDIA compute capabilities 8.6, 8.9, and 12.0. A 12 GB VRAM target does not guarantee that the longest context fits on every card. Close other GPU-heavy applications or reduce context when memory is tight.

The universal runtime is a release candidate. The runtime page lists exactly what was tested. Windows and unmeasured GPU/model combinations should not inherit Linux benchmark claims.

Measured speed

These are workload-specific measurements, not a controlled comparison between GPUs or a guaranteed speed for every prompt.

The RTX 3060 long-context test processed 92,879 input tokens plus 128 generated tokens and matched the recorded reference output. Prefill was 272.60 tokens/s and decode 21.06 tokens/s, with just 215 MiB minimum free VRAM. This checks a particular runtime workload; it is not a broad model-quality evaluation.

Quality scores from other checkpoints in the model family are not presented as an evaluation of this exact downloadable file. No external leaderboard result is claimed here.

File integrity

Download sizes use decimal GB. Model weights have not changed during the runtime documentation update.

Related checkpoint

L0xRE-27b-Low-MTP is a separate, larger checkpoint with an additional prediction block. Its file hash, runtime setup and qualification are distinct. Use the main model file on this page for the recommended drafter-based setup.

Model origin and license

L0xRE-27b-Low is a quantized derivative of Qwen3.8-27B, with low-bit projections and higher-precision embeddings/output tensors. The model's configured native context capacity is 262,144 tokens; supported and tested serving contexts are determined by the runtime profiles above.

Model weights remain under the base model's research/community terms: license. The L0xRE runtime tooling is Apache-2.0; applicable runtime and third-party notices ship with its packages. Model licensing and runtime licensing are separate.