Instructions to use RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4 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("RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4")

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 RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4 with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4"

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": "RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4 with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4 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 "RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4"

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 RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4 with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4"

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 "RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4" \
  --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"

LFM2.5 2.6B · Calibrated GPTQ MXFP4

An activation-calibrated, native-format MXFP4 conversion of LiquidAI/LFM2.5-2.6B. This is not a naive round-to-nearest repack: its 166 quantized linear layers were solved sequentially with activation Hessians and native E8M0 scale search.

Status: ready for MLX deployment and integrity-verified. AIME25, IFBench, LiveCodeBench v6, and every locally runnable BFCL v4 category have complete, receipted results. Multi-IF is being rerun with final-answer-only response handling and is not reported as a final score.

At a glance

Property	Value
Value format	MXFP4 E2M1
Scale format	E8M0, one scale per 32 weights
Solver	Sequential GPTQ with native-scale search and non-regression fallback
Quantized layers	166 linear layers
Preserved precision	Tied embedding/output head in BF16
Calibration	128 × 1,024-token disjoint windows; 131,072 tokens total
Damping	0.01
Packed biases	None

Held-out quality check

WikiText-2 test, fingerprint a46124b21ac53738, using 32 × 512-token predictions:

Checkpoint	Perplexity	Relative to BF16
BF16 source	48.3791	baseline
Native MXFP4 RTN	55.3957	+14.50%
GPTQ MXFP4	48.4551	+0.157%

GPTQ reduces perplexity by 12.529% versus the matched native MXFP4 RTN control while remaining within 0.157% of BF16 on this check.

Completed benchmark results

Benchmark	Scope	GPTQ MXFP4	Liquid published reference
AIME25	30 problems × 4 samples	26.67%	51.87%
IFBench	300 examples	47.33%	59.17%
BFCL v4	20 runnable categories; 4,906 scored examples	43.06%	56.88%
LiveCodeBench v6	1,055 problems	51.47%	59.41%

AIME25 and IFBench use the pinned Liquid OpenBench harness, temperature 0.6, 32,768 maximum output tokens, and matched local seed 42. Liquid did not publish its seed. IFBench evaluates the visible answer after the model's </think> boundary; the corrected scorer was checked by reproducing the original raw score exactly before applying that boundary. The BFCL score is the official partial-overall produced by the pinned BFCL harness; web_search_base and web_search_no_snippet are excluded because they require unavailable SerpAPI credentials. The Liquid BFCL reference includes those categories, so it is context rather than an exact matched control.

Runtime

The checkpoint uses native MLX block-scaled tensors and runs with current MLX and mlx-lm on Apple Silicon.

mlx_lm.generate \
  --model RESMP-DEV/LFM2.5-2.6B-GPTQ-MXFP4-calibrated \
  --prompt "Explain why calibrated quantization can outperform naive RTN."

For the CUDA quality campaign, the calibrated tensors were exactly dequantized, restored to Hugging Face LFM2 convolution layout, stored as BF16, and served by vLLM 0.26.0. This evaluates the calibrated checkpoint values under BF16 CUDA accumulation; it is not native packed-MXFP throughput or bit-identical MLX accumulation. A throughput smoke test on one RTX PRO 6000 Blackwell measured 1,410.6 aggregate output tokens/second at eight concurrent 64-token requests.

Reproducibility and integrity

Base-model revision: 403f92a2c4d78f2505f874b8b1b713cc0c9b9ae8
Calibration snapshot: private prerelease commit 2604ec2a4df3756c155648451364cd49c9dbb189; all eight non-card/config artifacts, including both weight shards and the tokenizer, are byte-identical to the public revision above
Calibration seed: 17
Calibration token SHA-256: b685dd74f7732611b78696f108ba50acbccb8905cb4f72ce1b49575b1b85858f
Layer-set SHA-256: cf1654096a8d86bb4dbaf37412adc6992a3767fe6b9114a788764ea4d98b2a65
model.safetensors SHA-256: a57eda09de8c9531007c2858d32086dd0cd641cc7bb94e02a3c746d8cc793329

Fresh loading and generation succeeded. All packed weights and scales were independently compared byte-for-byte with the CUDA calibration receipt. See GPTQ_RECEIPT.md for the compact provenance record.

Evaluation scope

Complete: AIME25, IFBench, LiveCodeBench v6, and the 20 locally available BFCL v4 categories. Re-evaluation in progress: Multi-IF with final-answer-only response handling. IFStruct and the remaining Liquid release rows are not yet complete. Disclosure- or credential-gated rows remain separate rather than being approximated silently.

Downloads last month: 124

Safetensors

Model size

3B params

Tensor type

BF16

U32

MLX

Hardware compatibility

4-bit

Model tree for RESMP-DEV/LFM2.5-2.6B-MLX-GPTQ-MXFP4

Base model

LiquidAI/LFM2.5-2.6B-Base

Finetuned

LiquidAI/LFM2.5-2.6B

Quantized

(119)

this model