Instructions to use Johnny5b/Ornith-9B-CORTEX-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Johnny5b/Ornith-9B-CORTEX-GGUF 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 Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M

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 Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M

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 Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M

Use Docker

docker model run hf.co/Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use Johnny5b/Ornith-9B-CORTEX-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Johnny5b/Ornith-9B-CORTEX-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Johnny5b/Ornith-9B-CORTEX-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M

Ollama
How to use Johnny5b/Ornith-9B-CORTEX-GGUF with Ollama:
```
ollama run hf.co/Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M
```
Unsloth Desktop

How to use Johnny5b/Ornith-9B-CORTEX-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M

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": "Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Johnny5b/Ornith-9B-CORTEX-GGUF with Docker Model Runner:
```
docker model run hf.co/Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M
```

Lemonade

How to use Johnny5b/Ornith-9B-CORTEX-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Ornith-9B-CORTEX-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use Johnny5b/Ornith-9B-CORTEX-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M

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 Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Johnny5b/Ornith-9B-CORTEX-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M

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 "Johnny5b/Ornith-9B-CORTEX-GGUF:Q4_K_M" \
  --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"

Ornith-9B-CORTEX-GGUF

Ornith-1.5-9B with the CORTEX protocol fine-tuned into the weights (QLoRA → merged → GGUF Q4_K_M), built per docs/CUSTOM-MODEL-BUILD-PLAN.md in Johnny5b/portable-ai-runner.

Files:

Ornith-9B-CORTEX-v3-Q4_K_M.gguf — current, recommended (5.63 GB, sha256 ca2803a2b9510e3d4cb5a51e5e244e8d616c5ec8d1df7193005364834a7d7b93). Passes the full behavior bar: model-emitted patch applied by the harness, no fallback, correct §S:REPAIR state (see evidence below). Trained on the v3 corpus: identity examples + in-scope Q&A for the expanded knowledge-vault domains + repair-patch completion.
Ornith-9B-CORTEX-v2-Q4_K_M.gguf — v2 (5.6 GB), kept for provenance. Passed the same behavior bar with the §S:AWAKE state nit.
Ornith-9B-CORTEX-Q4_K_M.gguf — v1 (5.6 GB), kept for provenance. Learned the trace header + RST scaffold but not patch-JSON emission.

Upstream: ornith-ai/Ornith-1.5-9B, MIT license. Attribution: Ornith-1.5 by ornith-ai.

Speed

GPU (A10G-large 24 GB, llama.cpp server-cuda, -ngl 99, ctx 2048, max_tokens 512):

Model	Gen tok/s	Job
Base Ornith-1.5-9B Q4_K_M	76.60 / 76.71	`6aad81ec51992417dfcc794a`
v1 CORTEX Q4_K_M	77.18 / 77.20	`6ab0994751992417dfcce80f`
v2 CORTEX Q4_K_M	77.19 / 77.16	`6ab0b80c52d0dbd7f1d773b1`

v3 CPU same-box A/B (cpu-upgrade, llama-bench b11175, -p 29 -n 64 -r 1; GPU flavors were unavailable at verify time, absolute numbers are contention-dominated — read the RELATIVE column):

Model	pp29 t/s	tg64 t/s	Job
v2 CORTEX Q4_K_M	3.01	0.27	`6ab5ee1652d0dbd7f1d8e12e`
v3 CORTEX Q4_K_M	1.95	0.30	`6ab5ee1652d0dbd7f1d8e12e`

Generation throughput parity confirmed (0.27 vs 0.30 — v3 ≥ v2). The pp29 gap is single-rep noise on a heavily contended box (the same box measured v2 at 2.58 pp29 an hour earlier, and v3 at 18.16 pp29 on its conversion load-gate). Structurally the files are identical for compute: same architecture, 427 tensors, same quant types, both 5.23 GiB. Behavioral corroboration: the v3 trial generated its full protocol answer in 411 s vs v2's 810.7 s on the same flavor.

Live trial evidence chain (`scripts/cortex_live_trial.py`, CPU)

Run	Gen cap	Patch extractor	Model output	Gate carried by	Job
v1 @ stock harness	300	stock	header + RST ✓, patch JSON never emitted	AST fallback	`6ab09ef851992417dfcce887`
v2 @ stock harness	300	stock	truncated preamble (budget)	AST fallback	`6ab0b80c51992417dfcceac4`
v2 corrected harness	800	fixed	full protocol + valid patch JSON	model patch ✓	`6ab0c48f51992417dfccebaf`
v3 corrected harness	800	fixed	full protocol, correct `§S:REPAIR`, valid patch JSON	model patch ✓	`6ab5e36a52d0dbd7f1d8df53`

The v3 run verbatim (2026-09-25):

[TRACE: §R:§EVO_AST:PYGGI | §T:repair_run | §S:REPAIR]
Diagnosis: the function divides by len(nums) with no empty-input guard, so avg([]) raises
ZeroDivisionError instead of ValueError.
{"find": "    return sum(nums) / len(nums)", "replace": "    if not nums:\n        raise ValueError('empty input')\n    return sum(nums) / len(nums)"}
===== REPAIR - model patch applied =====
===== REPAIR - pytest gate =====  3 passed in 0.01s
TRIAL PASSED   TRIAL_EXIT_CODE=0

v3 fixes v2's residual state nit (§S:REPAIR is now emitted, not §S:AWAKE) and completes the protocol in half the generation time (411 s vs 810.7 s).

The two harness fixes (required for the corrected-harness runs)

Generation cap 300 → 800 in chat() — the stock cap was sized for a non-reasoning 0.6B; a reasoning-style 9B spends >300 tokens thinking and the patch JSON never emits (measured, twice).
Patch extractor — parse the last line starting with {, no .replace("'", '"') quote mangling (it broke valid patches containing ValueError('empty input')).

A ready-to-run patcher that applies exactly these two fixes (byte-verified, with backup and audit-ledger line) ships at tools/apply_harness_fixes.py in Johnny5b/cortex-knowledge-vault.

Training

Adapter v3 (Johnny5b/Ornith-9B-CORTEX-LoRA-v3): same QLoRA recipe as v2 (r=16, α=32, TRL 1.13.0, 2 epochs, lr 2e-4 cosine, eff. batch 8, bf16/nf4) on the 404-example v3 corpus. 102/102 steps, final train_loss 0.434, token accuracy 0.917. Jobs: train 6ab355f252d0dbd7f1d825ad, merge/quant 6ab5db5e6b030d633f69024c, publish 6ab5f11052d0dbd7f1d8e1b4.
Adapters v2 / v1: see the LoRA-v2 / LoRA repos for their recipes (identical except corpus).
Data: Johnny5b/cortex-sft-300 — data/cortex_sft.jsonl (v1: 304), data/cortex_sft_v2.jsonl (v2: 356), data/cortex_sft_v3.jsonl (v3: 404 = v2 + identity block + vault-domain Q&A). Anti-contamination invariant held: no example contains the trial's avg fixture — the trial remains a transfer test.
Dashboard: ornith-9b-cortex-qlora-v3-trackio.

GGUF conversion note (qwen35 MTP metadata)

llama.cpp's converter derives qwen35.nextn_predict_layers from the HF config's mtp_num_hidden_layers, but the merged model carries no MTP tensors — the runtime then rejects the file. v3 (and future builds) convert with --no-mtp. The staged artifact additionally passed a load gate (llama-bench executed against the quantized file) before publishing, after the sha256 assertion. Conversion job: 6ab5db5e6b030d633f69024c.

Honesty ledger

Fine-tune deltas are measured ONLY by the tests above. Everything else is UNMEASURED — including the v3 identity examples ("I'm Ornith…"): trained, but no identity probe was run in the verification harness. The model may still self-identify as Qwen (its pretraining).
Both behavior passes required the two harness fixes; against the stock trial script the AST fallback still carries the gate. Harness limitation, documented, not hidden.
v3 residual nits: (a) thinking residue still leaks into the raw reply before the trace header at the 800-token budget; (b) the header's rule field rendered as §R:§EVO_AST:PYGGI (double-§ formatting quirk) though the state and patch were correct.
The v3 speed gate is a CPU same-box A/B, not the GPU number: GPU flavors were unavailable platform-wide at verify time. Relative parity only; absolute GPU tok/s for v3 is unmeasured.
Architecture correction to the source plan: Ornith-1.5-9B is a dense hybrid (3:1 linear:full attention, no MoE routers) — the "small active fraction" label in the source BENCHMARKS.md does not match this config.

Usage

llama.cpp / llama-cpp-python ≥ 0.3.35 (arch qwen35). Reasoning-style model: allow ≥800 tokens generation budget — at 300 the thinking budget eats the reply (measured, three times).

Downloads last month: 366

GGUF

Model size

9B params

Architecture

qwen35

Hardware compatibility

4-bit

Model tree for Johnny5b/Ornith-9B-CORTEX-GGUF

Base model

ornith-ai/Ornith-1.5-9B

Quantized

(97)

this model

Ornith-9B-CORTEX-GGUF

Speed

Live trial evidence chain (scripts/cortex_live_trial.py, CPU)

The two harness fixes (required for the corrected-harness runs)

Training

GGUF conversion note (qwen35 MTP metadata)

Honesty ledger

Usage

Model tree for Johnny5b/Ornith-9B-CORTEX-GGUF

Live trial evidence chain (`scripts/cortex_live_trial.py`, CPU)