Instructions to use YourHighnessLA/L0xRE-27b-Low with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use YourHighnessLA/L0xRE-27b-Low with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf YourHighnessLA/L0xRE-27b-Low:Q2_K
# Run inference directly in the terminal:
llama cli -hf YourHighnessLA/L0xRE-27b-Low:Q2_K

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf YourHighnessLA/L0xRE-27b-Low:Q2_K
# Run inference directly in the terminal:
llama cli -hf YourHighnessLA/L0xRE-27b-Low:Q2_K

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf YourHighnessLA/L0xRE-27b-Low:Q2_K
# Run inference directly in the terminal:
./llama-cli -hf YourHighnessLA/L0xRE-27b-Low:Q2_K

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf YourHighnessLA/L0xRE-27b-Low:Q2_K
# Run inference directly in the terminal:
./build/bin/llama-cli -hf YourHighnessLA/L0xRE-27b-Low:Q2_K

Use Docker

docker model run hf.co/YourHighnessLA/L0xRE-27b-Low:Q2_K

LM Studio
Jan

vLLM

How to use YourHighnessLA/L0xRE-27b-Low with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "YourHighnessLA/L0xRE-27b-Low"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "YourHighnessLA/L0xRE-27b-Low",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/YourHighnessLA/L0xRE-27b-Low:Q2_K

Ollama
How to use YourHighnessLA/L0xRE-27b-Low with Ollama:
```
ollama run hf.co/YourHighnessLA/L0xRE-27b-Low:Q2_K
```
Unsloth Desktop

How to use YourHighnessLA/L0xRE-27b-Low with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf YourHighnessLA/L0xRE-27b-Low:Q2_K

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "YourHighnessLA/L0xRE-27b-Low:Q2_K"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use YourHighnessLA/L0xRE-27b-Low with Docker Model Runner:
```
docker model run hf.co/YourHighnessLA/L0xRE-27b-Low:Q2_K
```

Lemonade

How to use YourHighnessLA/L0xRE-27b-Low with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull YourHighnessLA/L0xRE-27b-Low:Q2_K

Run and chat with the model

lemonade run user.L0xRE-27b-Low-Q2_K

List all available models

lemonade list

Hermes Agent

How to use YourHighnessLA/L0xRE-27b-Low with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf YourHighnessLA/L0xRE-27b-Low:Q2_K

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default YourHighnessLA/L0xRE-27b-Low:Q2_K

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use YourHighnessLA/L0xRE-27b-Low with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf YourHighnessLA/L0xRE-27b-Low:Q2_K

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "YourHighnessLA/L0xRE-27b-Low:Q2_K" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

L0xRE-27b-Low

A 27B hybrid model tuned to run well on consumer GPUs. This is the standard build: a low-bit quantized body with high-precision embeddings, designed to be paired with the DFlash2 drafter for speculative decoding. It runs a 12 GiB certified configuration on an RTX 4090 and a 16 GiB / 256K-context configuration on a 5090-class card.

Headline numbers (RTX 4090, L0xRE SM89 runtime, internal bench):

Config	Code decode	TTFT	VRAM
12 GiB / 80K ctx + DFlash2-Q4	115.4 ± 2.4 tok/s	289 ms	11.7 GiB
16 GiB / 256K ctx + DFlash2-Q4	116.3 ± 2.9 tok/s	303 ms	14.5 GiB
Full-GPU / 32K + DFlash2-Q4	139.5 ± 7.5 tok/s (peak 149.9)	292 ms	~20 GiB

Headline quality (internal suite, pass@1): 124–126 / 150 with thinking enabled (three independent passes), vs 117 / 150 for the native-quant baseline it was built from. See Quality for the full receipts table.

Which file to download

File	Use when
`L0xRE-27b-Low.gguf` (this repo)	Default. Pair with a DFlash2 drafter (included in this repo).
`L0xRE-27b-Low-MTP`	You want the model to self-accelerate with no drafter file (built-in MTP head).

Both files are the same base model and quantization. The MTP variant adds one nextn prediction block (+830 MB) used only for standalone self-drafting; with a DFlash2 drafter attached that block is idle, so the standard build is smaller and preferred.

Drafters (included in this repo)

Drafter	Certified config	Size
`Qwen3.8-27B-DFlash2-Q4_K_M.gguf`	both 12 GiB and 16 GiB configs (recommended)	1.14 GB
`Qwen3.8-27B-DFlash2-Q2_K.gguf`	fallback when VRAM is tight (16 GiB config: 107.4 vs 116.3 tok/s)	705 MB

Quickstart (SM89, RTX 4090-class — certified 12 GiB config)

llama-server -m L0xRE-27b-Low.gguf \
  -md Qwen3.8-27B-DFlash2-Q4_K_M.gguf \
  -c 81920 -b 1024 -ub 64 -np 1 -t 8 -ngl 99 -fa on \
  -ctk kvarn3 -ctv kvarn2 \
  --spec-type draft-dflash --spec-draft-n-max 3 --spec-draft-ngl 99 \
  --spec-draft-type-k q4_0 --spec-draft-type-v q4_0 \
  --jinja --reasoning on --reasoning-effort low --reasoning-budget 8192

16 GiB / 256K config: -c 262144 -ub 512 -ctv kvarn3, keep the Q4_K_M drafter with --spec-draft-type-k q4_0 --spec-draft-type-v q4_0, and set GGML_KVARN_WINDOW_CHUNK=16384.

Prebuilt runtime: release beellama-sm89-v0.4.7-r1 in L0xRE-BeeLLama-Low (./l0xre serve --profile 12gb -m L0xRE-27b-Low.gguf -md Qwen3.8-27B-DFlash2-Q4_K_M.gguf).

These require the L0xRE runtime (branch release/l0xre-sm89-v0.4.7 of L0xRE-BeeLLama-Low), not stock llama.cpp. SM120 (5090) users: release beellama-sm120-v0.4.7-r9 in the same repo. Catalog and ledger: seanyourhighness/L0xRE.

File identity

File	Bytes	SHA-256
`L0xRE-27b-Low.gguf`	8,619,127,680	`b0849250c633aa93853bf119a877dbdafdd7b1a4ebb672bd6b6a1439906f3543`
`Qwen3.8-27B-DFlash2-Q4_K_M.gguf`	1,143,006,816	`1a25c56858e1ebe93f2718ac1d49d1151f9323325c1bbfd6209370f4db131ebd`
`Qwen3.8-27B-DFlash2-Q2_K.gguf`	705,430,880	`e3eb7705404817cdbcdabe56049a1952b3b37bcc8df6e4d4efaec5d41563fb7e`

Architecture: 27B-parameter hybrid on the Qwen3.8-27B base — 64 transformer blocks; low-bit (≈2–3 bpw) projections, higher-precision embeddings/output, 262,144 native context, 248,320 vocab.

Qualification boundaries

Qualified on the L0xRE BeeLLama runtime only — not stock llama.cpp, Ollama, LM Studio, Transformers, or SGLang.
SM89 (4090-class): qualified at the two configs above, single slot. MTP-standalone, longer contexts, and multi-slot serving are not part of the qualification.
SM120 (5090): RTX 5090 / WSL, one-slot 8K and 32K profiles (beellama-sm120-v0.4.7-r9).
Speed vs matched native GGUF (SM120 r9 release): prefill 0.984–0.990, decode 0.963–0.981, DFlash2 prose 0.994, DFlash2 code 1.004.

Quality

Internal 8-pack suite (toolcall, instruct-follow, structured-output, data-extract, reason-math, bug-find, agent-20, cli-40 — 150 items per pass, pass@1):

Configuration	pass 1	pass 2	pass 3
L0xRE-27b-Low (thinking on)	124 / 150	121 / 150	126 / 150
L0xRE-27b-Low (thinking off)	116 / 150	—	—
Native-quant baseline (thinking off)	117 / 150	—	—

Run on z840 / RTX 4090 (SM89) via the BeeLLama server. The suite is internal and non-canonical — treat these as directional, not leaderboard results. Broader n/150 runs are scheduled and will be appended here as receipts land.

License

This is a quantized derivative of Qwen3.8-27B and is distributed under the base model's research/community terms (LICENSE); the L0xRE runtime tooling around it is Apache-2.0. Full per-component notices: seanyourhighness/L0xRE.

Provenance

Built from Qwen/Qwen3.8-27B with the E3 hybrid quant recipe (low-bit projections, high-precision embeddings). Performance figures above are from the L0xRE SM89 runtime with the champion decode kernels (port commit 107970f, receipts in bench-receipts/ of branch release/l0xre-sm89-v0.4.7).

Downloads last month: -

GGUF

Model size

5B params

Architecture

escha

Hardware compatibility

2-bit

4-bit

View +1 variant

Model tree for YourHighnessLA/L0xRE-27b-Low

Base model

Qwen/Qwen3.8-27B

Quantized

(1234)

this model