Instructions to use SkyIsNotGreen/hive-conductor-1.7b-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use SkyIsNotGreen/hive-conductor-1.7b-v1 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="SkyIsNotGreen/hive-conductor-1.7b-v1")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("SkyIsNotGreen/hive-conductor-1.7b-v1", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use SkyIsNotGreen/hive-conductor-1.7b-v1 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 SkyIsNotGreen/hive-conductor-1.7b-v1:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf SkyIsNotGreen/hive-conductor-1.7b-v1:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf SkyIsNotGreen/hive-conductor-1.7b-v1:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf SkyIsNotGreen/hive-conductor-1.7b-v1:Q4_K_M

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf SkyIsNotGreen/hive-conductor-1.7b-v1:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf SkyIsNotGreen/hive-conductor-1.7b-v1:Q4_K_M

Use Docker

docker model run hf.co/SkyIsNotGreen/hive-conductor-1.7b-v1:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use SkyIsNotGreen/hive-conductor-1.7b-v1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "SkyIsNotGreen/hive-conductor-1.7b-v1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SkyIsNotGreen/hive-conductor-1.7b-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/SkyIsNotGreen/hive-conductor-1.7b-v1:Q4_K_M

SGLang

How to use SkyIsNotGreen/hive-conductor-1.7b-v1 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "SkyIsNotGreen/hive-conductor-1.7b-v1" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SkyIsNotGreen/hive-conductor-1.7b-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "SkyIsNotGreen/hive-conductor-1.7b-v1" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SkyIsNotGreen/hive-conductor-1.7b-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use SkyIsNotGreen/hive-conductor-1.7b-v1 with Ollama:
```
ollama run hf.co/SkyIsNotGreen/hive-conductor-1.7b-v1:Q4_K_M
```
Unsloth Desktop

How to use SkyIsNotGreen/hive-conductor-1.7b-v1 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf SkyIsNotGreen/hive-conductor-1.7b-v1: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": "SkyIsNotGreen/hive-conductor-1.7b-v1:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use SkyIsNotGreen/hive-conductor-1.7b-v1 with Docker Model Runner:
```
docker model run hf.co/SkyIsNotGreen/hive-conductor-1.7b-v1:Q4_K_M
```

Lemonade

How to use SkyIsNotGreen/hive-conductor-1.7b-v1 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull SkyIsNotGreen/hive-conductor-1.7b-v1:Q4_K_M

Run and chat with the model

lemonade run user.hive-conductor-1.7b-v1-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use SkyIsNotGreen/hive-conductor-1.7b-v1 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf SkyIsNotGreen/hive-conductor-1.7b-v1: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 SkyIsNotGreen/hive-conductor-1.7b-v1:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use SkyIsNotGreen/hive-conductor-1.7b-v1 with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf SkyIsNotGreen/hive-conductor-1.7b-v1: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 "SkyIsNotGreen/hive-conductor-1.7b-v1: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"

Hive Conductor 1.7B — Controller + Planner Specialist

Hive Conductor 1.7B

The Hive 1.7B roster

This model is one seat in the five-model Hive 1.7B family. They are role specialists, not five general chat models:

model	seats
Hive Conductor 1.7B	controller + planner
Hive Reader 1.7B	mechanics + RAG
Hive Auditor 1.7B	placement + critic
Hive Dispatcher 1.7B	dispatcher
Hive Tool Manager 1.7B	agency

The family is served and routed by Hive.

Hive Conductor 1.7B is a multi-role specialist for the Hive local control plane, built from LiquidAI/LFM3-1.7B-Base. It covers the controller and planner seats: assignment, capacity, dependency, scheduling, ordering, and JSON plan contracts.

This is the current best Hive 1.7B checkpoint for the Conductor seat. It supersedes the controller-only controller-sft-base-v3 checkpoint (54/60 HF, 57/60 GGUF), improving controller to 59/60 on both HF and GGUF while adding planner at Scion-level quality.

What it is

A two-role LoRA trained on synthetic, checker-validated Hive controller and planner tasks.
A fast, small alternative to a 35B generalist for the constraint family: controller 0.25 s median through the Q4_K_M GGUF, planner 0.11 s.
A single specialist that holds both roles without controller regression; the shared scheduling/constraint logic transfers both ways.
A candidate for the live controller/planner seats, with Scion-35B as the fallback.

What it is not

Not a replacement for a 27B+ generalist.
Not trained for router, judge, face, or general chat.
Not proven on arbitrary scheduling instances; see Limitations.

Use cases

Local controller. Assign model roles, jobs or workers under time, capacity and dependency constraints, and return a JSON assignment.
Capacity-aware planning. Check whether a proposed assignment or schedule fits the available workers, and produce an ordered plan when it does not.
Safe migration and rollback plans. Turn a short operational goal into ordered steps plus the risks to check before executing it.
Build and release planning. Order build, test, documentation and release jobs with dependencies, and keep the plan within a deadline.
Preflight for a multi-model stack. Produce the controller/planner JSON that the dispatcher and auditor seats can validate before the harness launches anything.

Conductor is the controller/planner seat. It returns plans and assignments; it does not execute them or act as a general chat model.

Evaluation

Held-out suites, greedy decoding. Baselines are the controller-only specialist (this model's starting point) and Scion-35B.

Controller — 60 tasks (9 original + 51 generated)

model	overall	original 9	schedules	median
Scion-35B	50/60 (0.833)	9/9	19/25	3.96 s
controller-only v3	54/60 (0.900)	9/9	19/25	0.99 s
this model — HF	59/60 (0.983)	9/9	24/25	0.94 s
this model — GGUF Q4_K_M	59/60 (0.983)	9/9	24/25	0.25 s

Controller per-family: assignments 9/9, validation 13/13, exact JSON 4/4, choices 6/6, ordered 3/3, schedules 24/25.

Planner — 77 tasks (17 original + 60 generated)

model	overall	original 17	schedules	median
controller-only v3	30/77 (0.390)	8/17	—	0.98 s
Scion-35B (original 17 only)	—	15/17	—	—
this model — HF	69/77 (0.896)	16/17	12/12	0.32 s
this model — GGUF Q4_K_M	66/77 (0.857)	15/17	10/12	0.11 s

Planner per-family (HF): plans/JSON 29/33, number 19/23, ordered 9/9, schedules 12/12.

Reproduce

The held-out task files, deterministic checkers and endpoint evaluator are in the Hive harness. The JSON files under eval/ are the exact evaluation records: greedy decoding, response_format enabled for JSON-shaped tasks, and Q4_K_M for the GGUF row. Weight hashes are in SHA256SUMS.

Training

Base: LiquidAI/LFM3-1.7B-Base
Starting checkpoint: controller-sft-base-v3 (controller-only LoRA)
Method: TRL SFT + LoRA, r=16, alpha=32, dropout 0.05
Targets: q/k/v/o, in_proj/out_proj, w1/w2/w3
Data: 3,402 rows — 1,701 controller replay + 1,701 planner, both synthetic and checker-validated; no user data
Format: LFM3 chat template with enable_thinking=false; assistant-only loss
Schedule: 2 epochs, 426 steps, batch 4 × grad-accum 4, lr 1e-4, max length 512
Hardware: one 20 GB AMD card, ROCm 7.2, bf16
Peak VRAM: 13.99 GiB

Prompt format

Controller system prompt:

You are the Hive controller. Return only the requested JSON.

Planner system prompt:

You are the Hive planner. Return only the requested JSON.

Controller example:

Two workers: W1 is free for 20 minutes, W2 for 30 minutes.
Tasks: A takes 20 minutes, B takes 5, C takes 15.
Assign every task to a worker without exceeding their free time.
Return ONLY JSON: {"assignments": [{"task": "A", "worker": "W1"}]}.
Include all three tasks.

Planner example:

Pack items of size 7, 5, 4, 3, and 2 into bins of capacity 10.
What is the minimum number of bins? Answer with the number only.

Expected planner output may include a compact calculation ending with the answer; the checker reads the last number.

Usage

Merged transformers model:

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "SkyIsNotGreen/hive-conductor-1.7b-v1"
model = AutoModelForCausalLM.from_pretrained(
    repo, subfolder="model", dtype="bfloat16", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained(
    repo, subfolder="model", trust_remote_code=True)

The merged model is the supported checkpoint; this release does not ship a standalone adapter.

GGUF with llama.cpp:

llama-server -m gguf/hive-conductor-1.7b-v1-Q4_K_M.gguf \
  --alias conductor-1.7b --host 127.0.0.1 --port 8156 \
  -ngl 999 -c 8192 -np 2 --jinja --reasoning off --no-webui

--reasoning off is required for this LFM3 template; without it the GGUF answer shape diverges from the HF model.

Run with Hive

hive serve can fetch this GGUF and start it as a role tier; see the Hive install guide. The direct llama-server command above is the minimal portable path.

Limitations

One controller schedule (controller-229) and one planner critical-path task (planner-102) still fail; both are arithmetic/branch-choice cases.
Planner number arithmetic is 19/23; packing and critical-path reasoning are the remaining weak spots.
One seed, two epochs. The result is a strong signal, not a full sweep.
The model is trained for the Hive controller and planner contracts; other prompts are out of scope.
Use a deterministic assignment/schedule guard in production; a model output is not a verified plan.

License and attribution

This is a derivative of LiquidAI/LFM3-1.7B-Base, released under the LFM Open License v1.0 (see LICENSE). The base model is by Liquid AI. Commercial use by an entity above USD 10M annual revenue is not licensed under LFM 1.0; review the license before commercial deployment.

Modified files: merged weights and GGUF conversion. The base license text is unchanged. See NOTICE for attribution.

Version

hive-conductor-1.7b-v1 (built from multirole-cp-v1), 2026-10-31
Supersedes controller-sft-base-v3
Merged: model/
GGUF: gguf/hive-conductor-1.7b-v1-Q4_K_M.gguf
Evaluations: eval/eval-controller.json, eval/eval-controller-gguf.json, eval/eval-planner.json, eval/eval-planner-gguf.json

Downloads last month: 16

GGUF

Model size

2B params

Architecture

lfm3

Hardware compatibility

4-bit

Model tree for SkyIsNotGreen/hive-conductor-1.7b-v1

Base model

LiquidAI/LFM3-1.7B-Base

Quantized

(12)

this model

Collection including SkyIsNotGreen/hive-conductor-1.7b-v1

Hive series

Collection

The fine-tuned controllers, encoders and decision models that make the Hive harness work. • 6 items • Updated 1 day ago