LFM2.5-2.6B

2.70B parameters, text in, text out, 128K tokens of context. 1 provider serves it.

Verified 2026-10-25 Rank #258 in LLMs · score 85.6 method v1 Sources 4 Downloads 30d 99.8k · #315 of 619 tracked

Changed this week: 6 new quantized builds on the Hub (109 in all).2026-10-24

Can I run it?1.6GB at INT4Estimatefits a 16 GB cardmethodCommunity: Liquid AI LFM2.5-2.6B — The 2.6B Parameter Edge Model That Defies Scaling Laws: Architecture Revolution and Deployment Practice | HappyRock →

What does it cost?$0/1M infree tier ⓘLiquid2026-10-24

Can I use it?lfm1.0Custom license; read the repo LICENSE before commercial hostinglicence file2026-10-24

Do people use it?99.8kdownloads, 30 days#315 of 619 we track · below the median+13 likes, downloads −14% this week852 likesHugging Face2026-10-24

Worth reading

Picked by rule from 18 tracked pieces: the lab's own post, the most shared, a field report, the latest.

The lab's own post

LFM2.5-2.6B: On-Device Agents

LFM2.5-2.6B is Liquid AI's 2.6B dense model built for agentic workloads with a 131K context window and native tool calling.

docs.liquid.ai 2026-08-04 3 HN pts 161 words sells inference

Latest · 59 days ago

LFM2.5-2.6B Is the On-Device Agent Model Worth Testing - Beagle

heybeagle.com 2026-08-28

18 best of 18 tracked

The 4 picks above, plus these 14. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleLFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edgeliquid.ai · 2026-08-12 · official
ArticleLFM2.5-DSpark: Up to 3.2x Faster Inference from H100 to MacBliquid.ai · 2026-08-21 · 18 HN pts · official
ArticleLFM2.5-8B-A1B: An Even Better On-Device Mixture-of-Expertsliquid.ai · 2026-05-28 · official
ArticleLFM2.5-Encoders for Fast Long-Context Inference on CPUhuggingface.co · 2026-08-06 · 1 HN pts
ArticleAccomplishing More with LFM2.5-VL-DSpark2tinteractive.com · 2026-09-24
ArticleLFM2.5-350M: No Size Left Behindliquid.ai · 2026-04-01 · official
ArticleLFM2.5-VL-450M: Structured Visual Intelligence, Edge to Cloudliquid.ai · 2026-04-27 · official
ArticleLiquid AI LFM2.5-2.6B — The 2.6B Parameter Edge Model That Defies Scaling Laws: Architecture Revolution and Deployment Practice | HappyRockhappyrock.cloud · 2026-08-09
ArticleLFM2.5-2.6B-Base Explained: 34T Tokens, Zero Benchmarksorcarouter.ai · 2026-08-05
ArticleLFM2.5-2.6B: On-Device Agent Model (2026) | explainx.ai Blog | explainx.aiexplainx.ai · 2026-08-05
DiscussionLFM2.5-350Mtwitter.com · 2026-04-02
ArticleField card: LFM2.5-2.6Bgraphometer.ai · undated
ArticleDeploy local agents everywhere with LFM2.5-2.6Bwpnews.pro · undated
ArticleLFM2.5 - a hybrid architecture that's neither Transformer nor Mamba | lilting channellilting.ch · 2026-02-08

Show all 14

Where to run it

Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.

#	Cheapest live row	In /1M	Out /1M	Context	Quant	Seen
1	Liquid · free tier	$0	$0	64K	fp8	Oct 24

All 1 provider →

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

Served by 1 API provider; heavily run locally (109 quantized builds); a big base for builders (89 fine-tunes, adapters and merges).

As a service1Liquid
On own hardware109quantized builds
Built on8965 fine-tunes · 23 adapters · 1 merges
In apps15Hugging Face Spaces

Top Spaces running it: librarian-bots/dataset-tldr ♥ 7 · geocine/MiniMax-H3-Prompt-Enhancer-2.6B ♥ 4 · apathy-exe/LFM2.5-2.6B ♥ 4 · prodnobody/space ♥ 1

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 16 GB card at INT4

About 1.6 GB at INT4, 3.2 GB at FP8, 6.5 GB at BF16: weights only, so leave room for the KV cache.

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2026-07-28), amber triangles are 9 dated articles (each links to it), grey ticks are repo commits.

What people built on it

198 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

198 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

GGUF LiquidAI/LFM2.5-2.6B-GGUF for llama.cpp, Ollama, LM Studio 2.5M downloads MLX lmstudio-community/LFM2.5-2.6B-MLX-8bit for Apple silicon 17k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

What the card says

3 cited sentences

Summary generated, cited

LFM2.5-2.6B is a 2.7B parameter text-generation model from LiquidAI designed for on-device deployment.^[1]^[2] The model uses the Lfm2ForCausalLM architecture and supports a 131,072 token context length.^[3]^[4] It is released under the lfm1.0 license and is suitable for agentic workloads and tool use.^[5]^[6]

parameters human
“designed for on-device deployment”
architectures
“Context length: 131,072 tokens”
license
“We recommend using it for agentic workloads, tool use”

Generated 2026-10-07 15:05 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

6 use cases

What it is good at 6 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

22 fields

Listed in LLMs

Facts

LFM2.5-2.6B has 2.70B parameters. (safetensors metadata, observed 2026-10-24)
Context length is 131,072 tokens (128K). (config.json max_position_embeddings, observed 2026-10-24)
License: lfm1.0 (friction: review). Custom license; read the repo LICENSE before commercial hosting. (license file, observed 2026-10-24)
Created 2026-07-28, last modified 2026-10-23. (HF API, observed 2026-10-24)

Pipeline tag	text-generation
Library	transformers
Architectures	Lfm2ForCausalLM
Model type	lfm2
Dtype	bfloat16
Params by dtype	BF16: 2,697,198,592
Size band	M
MoE	no
Context length	131,072 (config.json max_position_embeddings)
Languages	ar, zh, en, fr, de, hi, id, it, ja, ko, pl, pt, ru, es, th, vi
Papers	arXiv:2511.23404
License	lfm1.0 (other)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Lfm2ForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Lfm2ForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for LFM2.5-2.6B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	6.5 GB · Single GPU (<24 GB)
FP8 (1 byte/param)	3.2 GB · Edge (<4 GB)
INT4 (0.5 bytes/param)	1.6 GB · Edge (<4 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

40 variants

Lineage

Fine-tuned from LiquidAI/LFM2.5-2.6B-Base (same organization) Family: LiquidAI/LFM2.5 198 derivatives on the Hub

Base model: LiquidAI/LFM2.5-2.6B-Base

quantized: 109finetune: 65adapter: 23merge: 1

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
LiquidAI/LFM2.5-2.6B-GGUF	quantized	GGUF	—	—	other	LiquidAI	391	2.5M
DavidAU/LFM2.5-2.6B-Qwen3.8-Turbo-Brilliance-Power-X12-NEO-MAX-GGUF	quantized	GGUF	—	—	apache-2.0	DavidAU	168	42.9k
LiquidAI/LFM2.5-2.6B-DSpark	finetune	—	328M	S	other	LiquidAI	38	35.8k
Abiray/LFM2.5-2.6B-Heretic-Abliterated-GGUF	quantized	GGUF	—	—	apache-2.0	Abiray	44	29.5k
PhillyLabs/LFM2.5-2.6B-UNCENSORED-ABLITERATED-PHILADELPHIA-CLASS	quantized	GGUF	2.70B	M	other	PhillyLabs	11	22.5k
SC117/LFM2.5-2.6B-Uncensored-GGUF	quantized	GGUF	—	—	other	SC117	19	21.7k
bartowski/LiquidAI_LFM2.5-2.6B-GGUF	quantized	GGUF	—	—	other	bartowski	17	17.9k
lmstudio-community/LFM2.5-2.6B-MLX-8bit	quantized	MLX	2.70B	M	other	lmstudio-community	2	17k
Arki05/LFM2.5-2.6B-GGUF-shards	quantized	GGUF	—	—	other	Arki05	0	15.9k
LiquidAI/LFM2.5-2.6B-ONNX	quantized	—	—	—	other	LiquidAI	23	11k
Akahsizrr/fuse-1-Lite	finetune	—	5.72B	M	apache-2.0	Akahsizrr	124	3.1k
LiquidAI/LFM2.5-2.6B-MLX-8bit	quantized	MLX	2.70B	M	other	LiquidAI	3	9k
mlx-community/LFM2.5-2.6B-4bit	quantized	-4bit	2.70B	M	other	mlx-community	5	8.7k
lmstudio-community/LFM2.5-2.6B-GGUF	quantized	GGUF	—	—	other	lmstudio-community	0	8.9k
nassimjp/LFM2.5-2.6B-Pashto-Zi	finetune	—	2.69B	M	apache-2.0	nassimjp	1	7.7k
mradermacher/LFM2.5-2.6B-i1-GGUF	quantized	GGUF	—	—	other	mradermacher	2	7.4k
noctrex/LFM2.5-2.6B-heretic-uncensored-GGUF	quantized	GGUF	—	—	—	noctrex	3	7.3k
LiquidAI/LFM2.5-2.6B-MLX-4bit	quantized	MLX	2.70B	M	other	LiquidAI	4	6.9k
reaperdoesntknow/LFM2.5-2.6B-CyberSec	quantized	GGUF	2.70B	M	apache-2.0	reaperdoesntknow	0	6.1k
infosave/LFM2.5-cmf	quantized	—	—	—	other	infosave	1	5.8k
Davd-b01/lfm2.5-2.6B-thinkingcap-distiller-v2	quantized	GGUF	2.70B	M	apache-2.0	Davd-b01	3	5.5k
zaakirio/LFM2.5-2.6B-Uncensored-GGUF	quantized	GGUF	—	—	other	zaakirio	2	5k
bunnycore/LMF-2.5-2B-Code-GGUF	quantized	GGUF	—	—	—	bunnycore	3	4.5k
lmstudio-community/LFM2.5-2.6B-MLX-4bit	quantized	MLX	2.70B	M	other	lmstudio-community	0	4.3k
SC117/LFM2.5-2.6B-Uncensored	finetune	—	2.70B	M	other	SC117	9	3.2k
Davd-b01/lfm2.5-2.6B-thinkingcap-distiller	quantized	GGUF	2.70B	M	apache-2.0	Davd-b01	4	3.3k
litert-community/LFM2.5-2.6B	quantized	—	—	—	other	litert-community	5	3.2k
Schnuckade/LFM-2.5-Coder-2.6B	quantized	GGUF	2.70B	M	apache-2.0	Schnuckade	4	3.1k
LiquidAI/LFM2.5-2.6B-MLX-bf16	finetune	MLX	2.70B	M	other	LiquidAI	7	2.7k
AyoubChLin/lfm2.5-2.6b-fable5-coding-agent	finetune	—	2.70B	M	—	AyoubChLin	11	2.4k
mlx-community/LFM2.5-2.6B-8bit	quantized	-8bit	2.70B	M	other	mlx-community	3	2.3k
mradermacher/LFM2.5-2.6B-GGUF	quantized	GGUF	—	—	other	mradermacher	0	2.3k
LiquidAI/LFM2.5-2.6B-MLX-mxfp4	quantized	MLX	2.70B	M	other	LiquidAI	4	2k
badtheorylabs/Macaw	finetune	—	2.70B	M	other	badtheorylabs	11	1.6k
heretic-org/LFM-2.5-2.6B-heretic	finetune	—	2.70B	M	other	heretic-org	9	1.7k
mlboydaisuke/LFM2.5-2.6B-CoreAI	quantized	—	—	—	other	mlboydaisuke	0	2k
lmstudio-community/LFM2.5-2.6B-MLX-6bit	quantized	MLX	2.70B	M	other	lmstudio-community	0	2k
LiquidAI/LFM2.5-2.6B-MLX-nvfp4	quantized	MLX	2.70B	M	other	LiquidAI	4	1.8k
atakankutalp/LFM2.5-2.6B-heretic	quantized	GGUF	2.70B	M	other	atakankutalp	0	1.9k
experimentalmachines/LFM2.5-2.6B-ExecuTorch	quantized	—	—	—	other	experimentalmachines	0	1.8k

Repository changelog

9 commits

changed in the last 7 daysRepo updated: 2026-10-24Repo commits: 2026-10-23Other: 2026-10-24all changes

Repository changelog 9 commits

Date	Commit	SHA
2026-10-23	Update README.md	b212e75
2026-08-24	New SGLang link	654f946
2026-08-24	Add DSpark	aa2553c
2026-08-20	Update README.md	a334ee7
2026-08-07	Update README.md	ab00687
2026-08-06	Update README.md	403f92a
2026-08-06	fix: Add `"do_sample": true,` since this LFM model uses `"temperature": 0.1` (#7)	8a7295f
2026-08-05	Update config.json	a4e00e8
2026-08-04	Initial commit	dca1825

All providers

1 provider · 1 endpoint

Where to run 1 providers · 1 endpoints

Hosted endpoints observed for LFM2.5-2.6B, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.00/M.

Provider	Model ref	Context	Quant	$/M in	$/M out	tok/s	TTFT	Tools	Status	Observed
Liquid openrouter	liquid/lfm-2.5-2.6b:free	64K	fp8	$0.00	$0.00	—	—	yes	live	2026-10-24

Articles and coverage

18 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official4

LFM2.5-DSpark: Up to 3.2x Faster Inference from H100 to MacB

liquid.ai2026-08-21officialabout the LFM2.5 familyHN 18 pts · 0 comments1,933 words

Liquid AI releases DSpark draft models for LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B, integrating speculative decoding via llama.cpp and SGLang. The draft models use 5 layers with a block size of 9 and total approximately ~300M parameters, achieving up to 3.18x GPU speedup and up to 2.87x on-device gains for the 2.6B variant. While the 1.2B and 2.6B models see consistent improvements, the 8B-A1B variant only achieves an 18% improvement on edge devices due to current MoE implementation constraints in llama.cpp's Metal backend.written, reviewed, cited

The 8B-A1B draft model adds 65.5M parameters for its Markov head compared to 33.6M in the 1.2B and 2.6B variants.
For the 1.2B-Instruct draft, validation loss consistently tracks acceptance rate improvements across all training epochs.
The 8B-A1B draft model achieves only an 18% improvement on edge devices because verifying k tokens activates more experts and weight traffic.

LFM2.5-2.6B: On-Device Agents

docs.liquid.ai2026-08-04officialHN 3 pts · 0 comments161 words

LFM2.5-2.6B is Liquid AI's 2.6B dense model built for agentic workloads with a 131K context window and native tool calling. It runs on edge devices and is trained to work reliably inside agent harnesses like Hermes Agent, OpenClaw, and Pi. The chat template opens every assistant turn with a <think> tag so responses contain an explicit chain of thought before the final answer. Users can run it via Transformers, llama.cpp, vLLM, or SGLang.written, reviewed, cited

LFM2.5-Encoders for Fast Long-Context Inference on CPU

huggingface.co2026-08-06authorityabout the LFM2.5 familyHN 1 pts · 0 comments1,103 words

Liquid AI releases LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, two new models designed for fast long-context inference on CPU. These encoders support an 8,192-token context and are about 3.7× faster than ModernBERT-base at long context lengths. The models are trained with a masked language objective that masks 30% of tokens during training to achieve strong performance on GLUE, SuperGLUE, and multilingual tasks.written, reviewed, cited

The 350M variant is the choice when accuracy matters most.
The 230M variant is the choice for tighter hardware or higher throughput.
A base encoder provides general-purpose representations rather than direct task outputs.

LFM2.5-2.6B: Deploy Agents Everywhere

liquid.ai2026-08-04officialHN 3 pts · 0 comments1,566 words

Liquid AI releases LFM2.5-2.6B, a 2.6B parameter agentic model trained on ~34T tokens and available on Hugging Face. It features a 128K vocabulary and context-extension phase, decoding 220 tokens/s on an M5 Max and 113 tokens/s on a Ryzen AI Max+ 395 while staying under 2.5 GB. The model reaches 15K output tokens per second on an H100 GPU.written, reviewed, cited

The model doubles its vocabulary to 128K by extending the existing tokenizer in place rather than retraining the model from scratch.
Training uses a dedicated 128K context-extension phase so the model can handle the long inputs that agentic workloads require.
The training pipeline separates model optimization, inference, and environment execution into distinct components using FSDP, SGLang, and verl.

Analysis6

LFM2.5-2.6B Is the On-Device Agent Model Worth Testing - Beagle

heybeagle.com2026-08-28

Liquid AI LFM2.5-2.6B — The 2.6B Parameter Edge Model That Defies Scaling Laws: Architecture Revolution and Deployment Practice | HappyRock

happyrock.cloud2026-08-09

LFM2.5-2.6B-Base Explained: 34T Tokens, Zero Benchmarks

orcarouter.ai2026-08-05

LFM2.5-2.6B: On-Device Agent Model (2026) | explainx.ai Blog | explainx.ai

explainx.ai2026-08-05

The Interesting Number Isn't the Price. It's That There Isn't One. | Laura Martel

laura-martel.com2026-08-04

Field card: LFM2.5-2.6B

graphometer.ai

All coverage tracked for this model (18) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
Field card: LFM2.5-2.6B	graphometer.ai	—	2026-10-07	exa
Deploy local agents everywhere with LFM2.5-2.6B	wpnews.pro	—	2026-10-07	exa
Accomplishing More with LFM2.5-VL-DSpark	2tinteractive.com	2026-09-24	2026-10-07	hn
LFM2.5-2.6B Is the On-Device Agent Model Worth Testing - Beagle	heybeagle.com	2026-08-28	2026-10-07	exa
LFM2.5-DSpark: Up to 3.2x Faster Inference from H100 to MacB	liquid.ai official	2026-08-21	2026-10-07	hn
LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge	liquid.ai official	2026-08-12	2026-10-07	hn
Liquid AI LFM2.5-2.6B — The 2.6B Parameter Edge Model That Defies Scaling Laws: Architecture Revolution and Deployment Practice \| HappyRock	happyrock.cloud	2026-08-09	2026-10-07	exa
LFM2.5-Encoders for Fast Long-Context Inference on CPU	huggingface.co	2026-08-06	2026-10-07	hn
LFM2.5-2.6B-Base Explained: 34T Tokens, Zero Benchmarks	orcarouter.ai	2026-08-05	2026-10-07	exa
LFM2.5-2.6B: On-Device Agent Model (2026) \| explainx.ai Blog \| explainx.ai	explainx.ai	2026-08-05	2026-10-07	exa
LFM2.5-2.6B: On-Device Agents	docs.liquid.ai official	2026-08-04	2026-10-07	hn
LFM2.5-2.6B: Deploy Agents Everywhere	liquid.ai official	2026-08-04	2026-10-07	hn
The Interesting Number Isn't the Price. It's That There Isn't One. \| Laura Martel	laura-martel.com	2026-08-04	2026-10-07	exa
LFM2.5-8B-A1B: An Even Better On-Device Mixture-of-Experts	liquid.ai official	2026-05-28	2026-10-07	hn
LFM2.5-VL-450M: Structured Visual Intelligence, Edge to Cloud	liquid.ai official	2026-04-27	2026-10-07	hn
LFM2.5-350M	twitter.com	2026-04-02	2026-10-07	hn
LFM2.5-350M: No Size Left Behind	liquid.ai official	2026-04-01	2026-10-07	hn
LFM2.5 - a hybrid architecture that's neither Transformer nor Mamba \| lilting channel	lilting.ch	2026-02-08	2026-10-07	exa

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

LFM2 Technical Report this model

Alexander Amini, Anna Banaszak, Harold Benoit +30 · 2025-11-28 · PDF · HF paper page (74 upvotes)

Similar models

3 models

Similar models

phi-2 · 2.78B
same size band (M), same task (text-generation), permissive license
gemma-2-2b-it · 2.61B
same size band (M), same task (text-generation)
gemma-2-2b · 2.61B
same size band (M), same task (text-generation)

Sources

4 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:37 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:05 UTC · Last verified 2026-10-25

Follow LFM2.5-2.6B

Watch this model

/feeds/models/LiquidAI--LFM2.5-2.6B.xml(RSS 2.0) · every change to LFM2.5-2.6B
/feeds/families/LiquidAI--LFM2.5.xml(RSS 2.0) · every change across the LFM2.5 family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.