Instructions to use summerAIO/diffusion-lfm2.5-24b-a2b-blocktune with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use summerAIO/diffusion-lfm2.5-24b-a2b-blocktune with Transformers:

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

pipe = pipeline("text-generation", model="summerAIO/diffusion-lfm2.5-24b-a2b-blocktune")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("summerAIO/diffusion-lfm2.5-24b-a2b-blocktune")
model = AutoModelForCausalLM.from_pretrained("summerAIO/diffusion-lfm2.5-24b-a2b-blocktune", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use summerAIO/diffusion-lfm2.5-24b-a2b-blocktune with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "summerAIO/diffusion-lfm2.5-24b-a2b-blocktune"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "summerAIO/diffusion-lfm2.5-24b-a2b-blocktune",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/summerAIO/diffusion-lfm2.5-24b-a2b-blocktune

SGLang

How to use summerAIO/diffusion-lfm2.5-24b-a2b-blocktune 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 "summerAIO/diffusion-lfm2.5-24b-a2b-blocktune" \
    --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": "summerAIO/diffusion-lfm2.5-24b-a2b-blocktune",
		"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 "summerAIO/diffusion-lfm2.5-24b-a2b-blocktune" \
        --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": "summerAIO/diffusion-lfm2.5-24b-a2b-blocktune",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use summerAIO/diffusion-lfm2.5-24b-a2b-blocktune with Docker Model Runner:
```
docker model run hf.co/summerAIO/diffusion-lfm2.5-24b-a2b-blocktune
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

DLFM · BlockTune

Diffusion LFM2.5 · 24B-A2B · Merged BF16

Block prediction, measured end to end.

Quick start · Results · Research report · 日本語

License — LFM Open License v1.0

This model is derived from LiquidAI's LFM2.5-24B-A2B and remains subject to the LFM Open License v1.0, licensed by Liquid AI, Inc. The complete, unmodified upstream license text is included as LICENSE and licenses/LFM-Open-License-v1.0.txt.

The license contains a commercial-use revenue threshold defined as US$10,000,000 annual revenue or more; Section 5 sets out the commercial-use limitation and the qualified non-profit exception. Read the original terms for the operative conditions. Redistribution must retain the license and applicable notices and identify modified files. See NOTICE for this release's modifications and license-notice.txt for retained model/data attribution.

LFM Open License v1.0 is not an Apache-2.0 or MIT license. This derivative does not replace it or grant rights beyond applicable upstream model and data terms.

A research checkpoint—with the evidence attached

DLFM · BlockTune is an experimental block-diffusion derivative of summerAIO/diffusion-lfm2.5-24b-a2b. It incorporates the final rank-8 attention LoRA from a short adaptation study, then packages the resulting weights as a complete, sharded BF16 checkpoint. No separate LoRA application or base-model download is required once this repository is downloaded.

The release includes the diffusion sampler, batched best-of-N selection, repetition controls, reproducible configuration, and both positive and negative results.

Research preview—not a production-quality coding model. BF16 merging did not meet the predeclared prediction-equivalence tolerance. The fastest decoding setting produces mostly incorrect answers. No tested configuration combined the quality gate with 1,000 returned token/s. Full-checkpoint byte/structure checks are documented separately from model-loading tests; a fresh end-to-end reload of this assembled full checkpoint has not been run.

At a glance

Component	Release
Backbone	LFM2.5-24B-A2B-derived hybrid convolution / attention MoE
Stored precision	BF16 weights; inherited FP32 buffers preserved
Adaptation	Rank 8, alpha 16, q/k/v/out projections in all 10 attention layers
Runtime	Explicit block-diffusion sampler; causal prefix, bidirectional canvas
Evaluated hardware	H100 80GB for adaptation; A100 80GB for decoding
Tested temperature	0.9
Weight format	13 Safetensors shards, approximately 47.94 GB
Recommended status	Experimental; no quality-approved throughput preset

Quick start

Use an 80GB-class CUDA GPU for the demonstrated BF16 setup. Download this repository, then run the included inference script from its folder:

hf download summerAIO/diffusion-lfm2.5-24b-a2b-blocktune \
  --local-dir ./dlfm-blocktune

Inside dlfm-blocktune:

uv run --python 3.11 --locked python inference.py \
  --model . \
  --preset balanced \
  --prompt "Write only Python code defining add(a,b) that returns their sum. No explanation."

The balanced preset uses block32 / 16 steps / top-k20 / best-of4 at temperature0.9. It selects candidates using syntax, the function name requested in the prompt, model likelihood and repetition—not reference answers or hidden tests. Optional code-prefix extraction is enabled by default. This was the best observed speed/accuracy compromise among the completed small-task comparisons, but it did not pass the overall quality gate.

Throughput diagnostic—not a quality recommendation:

uv run --python 3.11 --locked python inference.py \
  --model . --preset throughput --seed 900 \
  --prompt "Explain masked diffusion text generation and autoregressive text generation in detail. Write a coherent essay of at least 1000 words comparing their operation, advantages, and limitations."

This preset uses block1024 / 1 step / top-k50. It prints the output to stdout and the actual output length, elapsed time and throughput to stderr.

Python API

from inference import load_model
from selection_runtime import cfg, generate_selected

sampler, tokenizer = load_model(".")
settings = cfg(block=32, steps=16, best_of=4, top_k=20, repair=True)
text, metrics = generate_selected(
    sampler, tokenizer,
    "Write only Python code defining square(x) that returns x*x. No explanation.",
    settings, seed=2201,
)
print(text)
print(metrics)

The native Transformers model can load the weights, but its ordinary model.generate() is autoregressive. Use inference.py for the block-diffusion path measured in this report. The loader deliberately avoids the upstream custom AutoModel mapping and does not enable the numerically unverified prefix cache.

What the experiments show

1 · Adaptation improved a narrow block-prediction metric

On 32 held-out Python-code blocks, with 128 context tokens and a fully masked 64-token target, single-forward token accuracy increased from 6.15% to 8.69% after adaptation. Entire-block exact match stayed at 0/32.

These are pre-merge adapter-path results, not a claim that the assembled BF16 checkpoint reproduces that accuracy. The test set was reused across exploratory stages; near-duplicate contamination and prior training exposure were not ruled out.

2 · More candidate tokens do not mean faster useful answers

All rows below use temperature0.9. Task correctness is a small suite of 6 English fact/code tasks × 3 seeds; speed is measured separately on a long-form prompt.

Decoding setting	Returned token/s, median	Task correctness
block1024 / 1 step / top-k50	8,623.29	1/18 · 5.6%
Same block/steps + 32-token-window bigram ban, controlled sampler	3,771.05	0/18 · 0%
block32 / 16 steps / best-of4 / top-k20	17.89	13/18 · 72.2%
Same quality-oriented setting, best-of16	13.24	13/18 · 72.2%
Same quality-oriented setting, best-of64	4.52	13/18 · 72.2%
Same quality-oriented setting, best-of256	1.33	13/18 · 72.2%

Read the speed column carefully. Timings are CUDA-synchronized and count actual tokens emitted before EOS. Output lengths differ. Best-of-N timings include candidate generation, rejection sampling, ranking and post-processing. Discarded candidates are not counted as returned tokens. The controlled sampler has additional CPU work and a slightly different candidate distribution; its slowdown is not an isolated estimate of repetition-control cost.

The 8,623 token/s figure is a throughput result with poor correctness—not a claim of useful coding output at that speed. High distinct-n-gram rates can simply reflect random, incoherent text.

3 · The merge has a known numerical discrepancy

The measured BF16 merge had top-1 agreement of 82.81–96.88% across four probes, below the fixed 98% per-probe acceptance threshold. Maximum absolute logit deviation reached 2.5996. Saved merged weights reload byte-exactly, but that is a different check from agreement with the unmerged LoRA computation.

See merge_validation.json and the full report.

Repetition controls and best-of-N

uv run --python 3.11 --locked python inference.py \
  --model . --preset balanced \
  --block 32 --steps 16 --best-of 4 --top-k 20 \
  --window 128 --ngram 3 --penalty 1.1 \
  --prompt "Write only Python code defining reverse(xs) returning the list in reverse order. No explanation."

This is a custom-control example, not the measured best preset. Stronger penalties and harder bans were not consistently better. Constraints inspect resolved, contiguous left context; with multiple diffusion steps, they do not guarantee a globally repetition-free completed canvas across previously masked gaps.

For the balanced preset, --long continues selected blocks up to max(1024, block) tokens, respecting EOS. --max-new-tokens controls the throughput preset. The sampler is fixed at temperature0.9 to match the experiments.

Release contents

model-00001-of-00013.safetensors ... model-00013-of-00013.safetensors
model.safetensors.index.json     # complete checkpoint index
config.json + tokenizer files
diffusion.safetensors            # learned mask embedding
inference.py                    # full-checkpoint loader and CLI
diffusion_runtime.py             # evaluated reference sampler
selection_runtime.py             # window constraints and best-of-N
REPORT.md                       # consolidated research report
reports/                        # detailed reports and raw measurements
release_manifest.json           # source/output hashes and replaced tensor names
merge_validation.json           # pre/post merge checks
adapter_reference.safetensors   # provenance/reproduction only; already merged
licenses/ + license-notice.txt

Do not apply adapter_reference.safetensors again to these already-merged weights. It is included only to preserve the adaptation provenance.

Provenance and licensing

Immediate source: summerAIO/diffusion-lfm2.5-24b-a2b, pinned revision 9b7f240522cfdbebc65c955be567c1474c8bf399.
Backbone ancestry: LiquidAI/LFM2.5-24B-A2B, LFM Open License v1.0.
Additional adaptation data: iamtarun/python_code_instructions_18k_alpaca, pinned revision 7cae181e29701a8663a07a3ea43c8e105b663ba1.
The source checkpoint also declares Alpaca CC-BY-NC-4.0 and ARC CC-BY-SA-4.0 training ancestry. Attribution and notices are retained in license-notice.txt and licenses/.

This release does not grant unrestricted commercial rights or replace upstream model/data terms. The additional dataset card did not specify a license in the metadata inspected during preparation; review the relevant source terms.

日本語サマリ

ライセンスはLiquid AI, Inc.のLFM Open License v1.0です。 原文全文を LICENSEに同梱し、年商基準に関する商用利用条件、再配布・変更表示条件を原文どおり適用します。追加学習やマージによって別ライセンスへ置き換えていません。

最終LoRAを統合した、フル重み付きの研究用チェックポイントです。 README・推論コード・学習と速度探索のレポート・生の評価データを同梱しています。

追加学習でblock内token正解率は6.15%→8.69%に改善。ただしこの数値はマージ前の評価です。
速度優先設定は8,623 token/sですが、課題正解率は1/18。高品質な生成速度を意味しません。
品質寄りの暫定設定はblock32・16step・top-k20・best-of4。13/18正解、返却速度17.89 token/sでした。
16〜256候補へ増やしても同じ設定の正解率は伸びませんでした。
マージ前後の数値同等性は未達。品質と1000 token/sの両立も未達です。

詳しくは総合レポート、学習レポート、 A100探索レポートをご覧ください。

Downloads last month: 62

Safetensors

Model size

24B params

Tensor type

F32

BF16

Model tree for summerAIO/diffusion-lfm2.5-24b-a2b-blocktune

Base model

LiquidAI/LFM2.5-24B-A2B-Base

Finetuned

LiquidAI/LFM2.5-24B-A2B

Finetuned

summerAIO/diffusion-lfm2.5-24b-a2b

Finetuned

(1)

this model

Quantizations

1 model

summerAIO
/

diffusion-lfm2.5-24b-a2b-blocktune

DLFM · BlockTune

Diffusion LFM2.5 · 24B-A2B · Merged BF16

License — LFM Open License v1.0

A research checkpoint—with the evidence attached

At a glance

Quick start

Python API

What the experiments show

1 · Adaptation improved a narrow block-prediction metric

2 · More candidate tokens do not mean faster useful answers

3 · The merge has a known numerical discrepancy

Repetition controls and best-of-N

Release contents

Provenance and licensing

日本語サマリ

Model tree for summerAIO/diffusion-lfm2.5-24b-a2b-blocktune

Dataset used to train summerAIO/diffusion-lfm2.5-24b-a2b-blocktune