Instructions to use DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-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 DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-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 DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS
# Run inference directly in the terminal:
llama cli -hf DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS
# Run inference directly in the terminal:
llama cli -hf DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS

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 DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS
# Run inference directly in the terminal:
./llama-cli -hf DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS

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 DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS
# Run inference directly in the terminal:
./build/bin/llama-cli -hf DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS

Use Docker

docker model run hf.co/DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS

LM Studio
Jan

vLLM

How to use DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-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": "DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS

Ollama
How to use DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF with Ollama:
```
ollama run hf.co/DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS
```
Unsloth Desktop

How to use DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS

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": "DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF with Docker Model Runner:
```
docker model run hf.co/DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS
```

Lemonade

How to use DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS

Run and chat with the model

lemonade run user.LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF-IQ4_XS

List all available models

lemonade list

Hermes Agent

How to use DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-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 DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS

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 DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS

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 "DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF:IQ4_XS" \
  --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"

benchmarks, comparable to the other ones I've given for your qwen fusion models

by pkp24 - opened 3 days ago

Discussion

pkp24

3 days ago

•

edited 3 days ago

I stopped the benchmarking because I didn't want to dedicate a huge amount more time to it, testing every single thinking mode was a ton of testing.

tldr nothing set for reasoning effort was by far the best compared to anything else.

Reasoning-mode benchmark: LFM2.5-2.6B Turbo-Brilliance NEO-MAX Q8_0

Tested model: DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF, file LFM2.5-2.6B-Q3.8-TBrilliance-NEO-MAX-Q8_0.gguf.

Snapshot taken 2026-09-26, 09:23. Run 1 is complete. Run 2 was still in progress: easy, hard and ultrahard were done for every mode, and abyssal was still running.

Setup

What changes between modes: only the reasoning mode. Every mode uses the same weights and the same llama.cpp server layout (4 slots × 128K context, f16 KV cache). The mode is set server-side with --reasoning-effort <mode>, which chooses the system block the chat template injects.
Baselines: off is stock LFM2.5 thinking with nothing injected. high is the template's default mode.
Sampling: thinking on for every test, with the model card's tester settings: temp 1.0, top-k 64, min-p 0.05, top-p 0.95, repeat penalty 1.0.
Limits: no token cap, so a response can run to the full 128K context. The only other stops are a 1800 s timeout per request and a repetition (loop) detector.
Test suites: four difficulty tiers (easy, hard, ultrahard, abyssal). The categories are coding (the program is run against tests), logic, JSON, distill, constraint, simulation, cx and flawstep. Planning tests are recorded but not scored.

How to read the scores:

Each test is a single sample at temp 1.0. A single suite can swing 10–20 points from one run to the next, so small gaps are not meaningful yet.
trunc, loop and blank each score 0 on that test:
- trunc: the response ran out of context.
- loop: the repetition detector abandoned it.
- blank: it finished without an answer.
wall min is not a speed comparison. Two modes ran at the same time on two servers sharing one GPU, so a mode's time depends on what the other server was doing.

Run 1 (complete: easy, hard, ultrahard, abyssal)

Scores (%)

mode	easy	hard	ultrahard	abyssal	mean	vs off	vs high	tok/test	loop	blank	wall min
off	95.0	75.0	15.8	1.2	46.8	+0.0	+9.1	11,594	0	5	37
deeptree	91.0	78.3	15.1	0.6	46.3	-0.5	+8.6	12,017	0	3	38
omni	93.8	64.6	14.8	3.1	44.1	-2.7	+6.4	11,287	0	3	33
einstein	93.8	65.0	16.7	0.5	44.0	-2.8	+6.4	13,393	0	0	43
spoon	68.8	65.0	23.8	0.5	39.5	-7.2	+1.9	11,801	2	0	28
low	83.8	55.4	12.5	4.1	39.0	-7.8	+1.3	12,454	0	3	39
socrates	70.0	73.8	8.3	3.7	38.9	-7.8	+1.3	12,350	0	2	44
high	73.8	60.8	12.4	3.5	37.6	-9.1	+0.0	13,835	0	0	43
hyper	67.5	57.8	16.4	0.6	35.6	-11.2	-2.0	12,438	0	1	38
logic	68.8	42.1	29.6	0.2	35.1	-11.6	-2.5	11,909	0	3	38
ultra	73.8	47.9	15.8	0.2	34.4	-12.3	-3.2	14,049	0	3	52
medium-low	78.8	41.7	13.9	3.2	34.4	-12.4	-3.2	11,652	1	0	42
medium	68.8	42.5	18.9	3.8	33.5	-13.3	-4.1	12,344	0	1	37

By category (all four suites pooled, %)

mode	coding	constraint	distill	json	logic	sim
off	57.1	4.2	25.0	50.0	62.5	12.5
deeptree	57.9	0.0	32.5	50.0	56.2	4.2
omni	62.5	0.0	29.2	50.0	43.8	12.5
einstein	46.6	4.2	27.5	50.0	56.2	12.5
spoon	24.0	0.0	27.5	58.3	50.0	25.0
low	27.3	12.5	32.5	50.0	50.0	1.8
socrates	49.2	0.0	30.0	41.7	43.8	12.5
high	20.2	3.1	27.5	50.0	56.2	12.5
hyper	30.7	8.3	25.0	50.0	37.5	12.5
logic	12.2	12.5	27.5	58.3	37.5	25.0
ultra	18.6	0.0	25.0	58.3	43.8	0.0
medium-low	21.1	4.2	25.0	50.0	43.8	12.5
medium	9.2	0.0	30.0	50.0	50.0	12.5

Where the modes disagree

Of 72 graded tests, 8 passed in every mode and 24 failed in every mode. 31 split, meaning some mode scored at least 50 points above another on that test.

Number of split tests on which each mode had the top score (ties count for every tied mode): off 20, omni 18, deeptree 17, einstein 17, socrates 15, spoon 14, low 13, high 13, medium-low 11, hyper 11, ultra 10, logic 10, medium 9.

Per-test scores on the 31 split tests (score ×100, blank = 0)

test	off	low	medium-low	medium	high	ultra	omni	deeptree	hyper	socrates	logic	einstein	spoon
abyssal:ab_chain3		100
abyssal:ab_chain4			100
abyssal:ab_chain5				100	100					100
easy:code_eval		100	100				100	100		100		100
easy:code_lvp	100		100		100		100	88	100	100		100
easy:code_minwin	100	100					100	100		17		100
easy:code_sliding	100						100	100		100		100
easy:code_wordbreak	100	100				100	100	100				100
easy:json_primes	100	100	100	100	100	100	100	100	100		100	100	100
easy:json_tool	100	100	100	100	100	100	100		100	100	100	100	100
easy:logic_div	100							100
easy:logic_race	100	100	100	100	100	100	100	100			100	100	100
hard:h_coin	100				100			100	100	100
hard:h_countsmaller	100		100				100	100	100	100			100
hard:h_domino	100	100		100	100	100	100	100	100	100	100	100	100
hard:h_edit	100	100				100	100	100		100		100	100
hard:h_lis	100						100	100	100	100
hard:h_lockers	100			100	100	100	100	100	100	100	100	100	100
hard:h_modexp	100	100	100		100	100	100	100	100			100	100
hard:h_regex	100						100		38	100	50	100
hard:h_sieve	100	100	100		100			100		100		100	100
hard:h_trap	100				100		100	100		100		100	100
ultrahard:uh_cover			100	100		100	100		100		59		100
ultrahard:uh_cs1	33	100	33						67		100	33
ultrahard:uh_j2						100		100			100		100
ultrahard:uh_mul	100			100								100
ultrahard:uh_parse	39				72	39	50	33	11		11		39
ultrahard:uh_sim15	100			100	100		100	33			100	100
ultrahard:uh_sim250													100
ultrahard:uh_sim60			100						100	100	100		100
ultrahard:uh_window		100		54		8		62	15		100	62

Run 2 (partial: easy, hard and ultrahard done for every mode; abyssal still running)

The mean below covers easy, hard and ultrahard only, so it is not comparable to run 1's four-suite mean.

Scores (%)

mode	easy	hard	ultrahard	mean	vs off	vs high	tok/test	blank	wall min
off	93.8	66.7	28.3	62.9	+0.0	+14.0	6,766	1	15
deeptree	95.0	75.0	12.8	60.9	-2.0	+12.1	9,083	0	19
omni	88.8	70.8	18.4	59.3	-3.6	+10.5	8,614	2	18
einstein	88.8	52.5	16.6	52.6	-10.3	+3.8	8,718	0	16
socrates	88.0	51.0	14.1	51.1	-11.8	+2.2	7,524	1	17
hyper	72.5	66.7	13.4	50.9	-12.0	+2.0	9,289	0	19
low	68.8	61.7	19.5	50.0	-12.9	+1.1	6,887	0	13
logic	73.8	53.5	19.4	48.9	-14.0	+0.1	8,195	1	15
high	75.0	61.3	10.3	48.8	-14.0	+0.0	8,769	1	19
medium-low	67.5	61.7	17.1	48.8	-14.1	-0.1	8,495	0	21
ultra	73.8	53.3	10.4	45.8	-17.1	-3.0	7,646	0	16
medium	61.3	54.6	18.1	44.6	-18.3	-4.2	8,989	3	20
spoon	78.8	35.4	12.3	42.2	-20.7	-6.7	8,320	0	11

By category (easy + hard + ultrahard pooled, %)

mode	coding	constraint	distill	json	logic	sim
off	77.5	8.3	33.3	77.8	66.7	50.0
deeptree	73.8	0.0	33.3	66.7	75.0	25.0
omni	69.5	25.0	33.3	66.7	66.7	25.0
einstein	59.3	6.2	36.7	55.6	66.7	8.3
socrates	64.5	0.0	33.3	55.6	58.3	8.3
hyper	54.8	0.0	33.3	66.7	58.3	0.0
low	15.6	0.0	43.3	77.8	66.7	25.0
logic	25.0	8.3	36.7	77.8	58.3	25.0
high	29.8	0.0	40.0	66.7	66.7	0.0
medium-low	18.5	8.3	43.3	66.7	58.3	50.0
ultra	21.2	0.0	43.3	77.8	50.0	0.0
medium	26.7	16.7	36.7	77.8	41.7	8.3
spoon	17.5	8.3	33.3	55.6	58.3	25.0

Where the modes disagree

Of 47 graded tests, 7 passed in every mode, 9 failed in every mode and 28 split.

Number of split tests on which each mode had the top score: off 22, deeptree 21, omni 20, hyper 14, socrates 14, einstein 14, low 12, high 12, logic 12, medium-low 11, medium 10, ultra 10, spoon 9.

Per-test scores on the 28 split tests (score ×100, blank = 0)

test	off	low	medium-low	medium	high	ultra	omni	deeptree	hyper	socrates	logic	einstein	spoon
easy:code_eval	100			100		100	100	100	100	86		100
easy:code_lvp	100		100				100	100		100
easy:code_minwin	100							100	100			100
easy:code_sliding	100						100	100		100	100	100	100
easy:code_wordbreak	100						100			100		100	100
easy:logic_div					100			100
easy:logic_fact	100	100	100		100	100	100	100		100	100	100	100
easy:logic_hh	100	100			100	100	100	100	100	100	100	100	100
hard:h_coin	100						100	100	100	100		100
hard:h_countsmaller	100		100	100	100		100	100	100			100
hard:h_domino	100	100	100	100	100		100	100	100	100	100	100
hard:h_edit	100	100					100	100		100
hard:h_graph	100	100	100	100	100	100	100	100	100		100	100
hard:h_lis	100							100		100
hard:h_modexp	100	100	100	100	100		100	100	100	100	100	100	100
hard:h_regex	50			100	100	100	100	100	100	75	75	100
hard:h_sieve		100	100			100	100	100	100				100
hard:h_toolevent	100	100	100	100	100	100	100	100	100	100	100		100
hard:h_trap	100				100	100	100	100	100		100	50
ultrahard:uh_cover	100	100		100	47			100	100	100		24	24
ultrahard:uh_crt	100	100	100	100	100	100	100	100	100	100		100	100
ultrahard:uh_cs1	33		33	67			100				33		33
ultrahard:uh_j2	100	100		100		100					100
ultrahard:uh_mul	100										100	100
ultrahard:uh_parse	6	28				11	50		39			11	39
ultrahard:uh_sim15	100	100	100	33			100			33	100	33	100
ultrahard:uh_sim60	100		100					100
ultrahard:uh_window			77		100		92		77	100	100	62

Observations so far

No mode beat stock thinking. off (no injection) led both runs, with deeptree and omni close behind. The template default high sat mid-pack, 9–14 points behind off.
Coding drives most of the differences. The coding scores span roughly 9–78% across modes, while distill and JSON barely move.
Ultrahard and abyssal are beyond this model. It scores about 8–30% on ultrahard and 0–4% on abyssal in every mode, so those suites say little about which mode is better.

DavidAU

Owner 2 days ago

•

edited 2 days ago

Thank you for detailed testing and notes.
These are very helpful.

A few notes:

2.6B is a limiting factor ; that being said LiquidAI did an exceptional job here however.
In run #1, on very hard/abyssal some of the reasoning modes clearly pulled ahead of "off" / "standard" - some by a long shot. ; I don't know if I would "mean" these kinds of numbers.
Without "gating" the specialized reasoning modes will attempt to solve problems they were not designed for. (this matches our own internal testing).
Gating would select the best mode automatically during testing, rather than "manually" as is the case now; IE you could run one test, for all modes VS 12.
The specialized reasoning is use case specific ; this is part explains the low scores. Larger models will disallow the "reasoning mode" automatically based on context ; if it doesn't match -> standard.
The reasoning modes were NOT calibrated for this specific model, they are "general", especially the advanced modes. IE min: 9B/27B sized models. (or larger)
An odd issue: sometimes the model will override/disagree with the reasoning mode, and cause issues itself.
Parameters for testing (ours) differ greatly from LiquidAi's published ones. (temp is especially critical, rep pen too -> especially as LFM suggests 1.1 vs 1 [off] from us)
Testing smaller para models is always more difficult // less range (our experience here) // likewise RAISING benchmarks (thru tuning) is also a lot harder too.
It appears your tests are primarily coding, coding related, and logic ; the alignment for all modes was "general" - with aligned reasoning modes the scores will likely reflect a change.

Again ; thank for publishing - this will help with refinements to reasoning, reasoning modes and other planned improvements.

pkp24

2 days ago

Yeah I generally use it for a lot of logic chains, tool usage, and coding so that's where most of the benchmarks are at.

"In run #1, on very hard/abyssal some of the reasoning modes clearly pulled ahead of "off" / "standard" - some by a long shot. ; I don't know if I would "mean" these kinds of numbers."
Due to how low the passing scores were with some of the much harder tests I don't know how well it really would do when run over and over again. sometimes models are on the cusp of failing or passing a test so it may flip between passing/failing multiple times if I run it multiple times. Since I only ran the full suite 1.2 times idk that the barely passing numbers mean too much.

"The specialized reasoning is use case specific ; this is part explains the low scores. Larger models will disallow the "reasoning mode" automatically based on context ; if it doesn't match -> standard."
If you can get auto gating somehow that'd certainly be far far easier to test with, and I could more reasonably run a full 10 runs as it'd be ~10 hours vs 72+ hours.

DavidAU

Owner 2 days ago

Note:
Another user alerted us to some jinja template bugs.
These have been corrected, and model is being retested.

Likely this may have affected some results/benches ; performance changes were noted after fixes -> ie net # of tokens as one direct result.

DavidAU

Owner 1 day ago

Revised ggufs now up.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment