Instructions to use sagaya/LFM2.5-1.2B-Instruct-control-mlx with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use sagaya/LFM2.5-1.2B-Instruct-control-mlx with MLX:

# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("sagaya/LFM2.5-1.2B-Instruct-control-mlx")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use sagaya/LFM2.5-1.2B-Instruct-control-mlx with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "sagaya/LFM2.5-1.2B-Instruct-control-mlx"

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "sagaya/LFM2.5-1.2B-Instruct-control-mlx"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use sagaya/LFM2.5-1.2B-Instruct-control-mlx with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "sagaya/LFM2.5-1.2B-Instruct-control-mlx"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "sagaya/LFM2.5-1.2B-Instruct-control-mlx"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "sagaya/LFM2.5-1.2B-Instruct-control-mlx",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use sagaya/LFM2.5-1.2B-Instruct-control-mlx with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "sagaya/LFM2.5-1.2B-Instruct-control-mlx"

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 sagaya/LFM2.5-1.2B-Instruct-control-mlx

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use sagaya/LFM2.5-1.2B-Instruct-control-mlx with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "sagaya/LFM2.5-1.2B-Instruct-control-mlx"

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 "sagaya/LFM2.5-1.2B-Instruct-control-mlx" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

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

sagaya/LFM2.5-1.2B-Instruct-control-mlx

An MLX build of LiquidAI/LFM2.5-1.2B-Instruct made small enough for 8 GB iPhones, with recipe 6bit-g64 (Plain 6-bit, groups of 64).

This is a modified version (fine-tuned for Control by distillation, then quantized) of the original model, distributed under the LFM Open License v1.0 (LICENSE). Commercial use by an organization with annual revenue of $10M or more is not licensed under it.

Quality against the full-precision model

Measure	This build	Full precision
Held-out perplexity	23.775	23.607
KL divergence on chat replies	0.0047 nats	0
Same next token on chat replies	97.7%	100 %
Recall of facts ~2,000 tokens back	4/4	4/4
Size	0.95 GB

Recipes we tried

Recipe	Size	Bits/weight	Perplexity	KL (chat)	Top-1 (chat)	Recall	Swift loads	Stops	Result
shipped-4bit	0.66 GB	4.5	29.49 (+24.9 %)	0.088	88.1%	4/4	Yes	–	Shipped today
shipped-6bit	0.95 GB	6.5	24.73 (+4.8 %)	0.020	96.0%	4/4	Yes	100%	Shipped today (default)
6bit-g64	0.95 GB	6.5	23.78 (+0.7 %)	0.005	97.7%	4/4	Yes	100%	Winner

Parity: our Swift engine vs Python mlx-lm

Check	Value	Result
6bit-g64 · top-1 agreement	100.0%	passed
6bit-g64 · greedy tokens	identical	passed
6bit-g64 · max \|Δlogit\|	0.219	passed
6bit-g64 · long-context recall	Swift 4/4 · Python 4/4	passed
6bit-g64 · chat template	same ids	passed
6bit-g64 · app engine reply	same opening	passed

Bench (Mac, the app's engine)

Check	Value	Result
6bit-g64 · stop reliability	18/18 ended on their own	passed
6bit-g64 · Mac decode speed	279 tok/s	info
6bit-g64 · first token (Mac)	0.02 s	info
6bit-g64 · peak memory (Mac)	1.00 GB	info
6bit-g64 · iPhone estimate	41.7 tok/s, cool	info

Evals

Task	shipped-6bit · Student before	untuned · Full precision	teacher · Teacher	6bit-g64 · Student after
ARC-Challenge (100)	45.0% ± 10%	45.0% ± 10%	44.0% ± 10%	40.0% ± 10%
GSM8K (50)	70.0% ± 13%	66.0% ± 13%	96.0% ± 5%	66.0% ± 13%
IFEval strict (50)	76.0% ± 12%	80.0% ± 11%	90.0% ± 8%	80.0% ± 11%
MMLU-Pro (3 per subject)	33.3% ± 15%	38.1% ± 13%	73.8% ± 13%	31.0% ± 14%
General average	56.1%	57.3%	76.0%	54.2%
Control tasks (quick)	13/65	9/65	54/65	26/65
actions	1/5	0/5	4/5	5/5
automations	0/7	0/7	5/7	0/7
calendar	0/12	0/12	10/12	5/12
contacts	0/3	0/3	3/3	0/3
general	2/3	2/3	1/3	3/3
health	0/2	0/2	2/2	0/2
longdoc	1/4	1/4	4/4	1/4
memory	2/5	2/5	4/5	2/5
permissions	1/5	0/5	5/5	2/5
places	0/1	0/1	1/1	1/1
recall	5/5	4/5	5/5	5/5
reminders	0/5	0/5	5/5	1/5
shortcuts	1/3	0/3	1/3	0/3
weather	0/2	0/2	2/2	1/2
web	0/3	0/3	2/3	0/3

Retains the original: 6bit-g64

Check	Original	This build	Same
Chat template	5,487 characters	5,487 characters	yes
Thinking switch	never	never	yes
Tool-call format	LFM2 (<\|tool_call_start\|> … pythonic calls)	LFM2 (<\|tool_call_start\|> … pythonic calls)	yes
Tokenizer (tokenizer.json)	present	same	yes
Tokenizer settings	tokenizer_config.json	same	yes
Special tokens	bos_token <\|startoftext\|> (1), eos_token <\|im_end\|> (7), pad_token <\|pad\|> (0)	same	yes
EOS and special token ids	bos_token_id 1, eos_token_id 7, generation bos_token_id 1, generation eos_token_id 7, generation pad_token_id 0, pad_token_id 0	same	yes
Generation settings	generation_config.json	same	yes

Distillation: held-out split

Measure	Before	After
Loss of the teacher's answers	1.093	0.525
Call a tool or answer, like the teacher	38.6%	84.1%
Same tool as the teacher	0.0%	59.3%
Same argument values	0.0%	30.9%

Use with mlx-lm

pip install mlx-lm
mlx_lm.generate --model sagaya/LFM2.5-1.2B-Instruct-control-mlx --prompt "Hello"

How it was tuned

Distilled for Control from Qwen3 30B-A3B Instruct (mlx-community/Qwen3-30B-A3B-Instruct-2507-4bit, license apache-2.0): 819 examples of Control-style requests (the app's tool specs and prompt format), LoRA, then quantized as above.

About 45% of the training rows are the original model's own answers (rehearsal), so it keeps what it knew before.

Prompts from databricks/databricks-dolly-15k, revision bdd27f4d94b9, license cc-by-sa-3.0, used under its terms with this credit.

Downloads last month: 149

Safetensors

Model size

1B params

Tensor type

U32

BF16

MLX

Hardware compatibility

6-bit

Model tree for sagaya/LFM2.5-1.2B-Instruct-control-mlx

Base model

LiquidAI/LFM2.5-1.2B-Base

Finetuned

LiquidAI/LFM2.5-1.2B-Instruct

Quantized

(126)

this model