Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › LFM2.5 family

LFM2.5-350M

Hugging Face ↗ Follow changes

LiquidAI lfm1.0 out 7 months ago

354M parameters, text in, text out, 125K tokens of context.

Verified 2026-10-25 Rank #511 in LLMs · score 76.2 method v1 Sources 3 Downloads 30d 74.4k · #349 of 619 tracked

Changed this week: 1 new quantized build on the Hub (87 in all).2026-10-24 · 2 new quantized builds on the Hub (88 in all).2026-10-24

Can I run it?0.2GB at INT4Estimatefits a 16 GB cardmethod
Can I use it?lfm1.0Custom license; read the repo LICENSE before commercial hostinglicence file2026-10-24
Do people use it?74.4kdownloads, 30 days#349 of 619 we track · below the median+0 likes, downloads +39% this week437 likesHugging Face2026-10-24

Worth reading

Picked by rule from 10 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Latest · 2 months ago

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

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 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.
liquid.ai 2026-08-21 18 HN pts 1.9k words about the LFM2.5 family family

10 best of 10 tracked

The 2 picks above, plus these 8. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. ArticleLFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edgeliquid.ai · 2026-08-12 · official
  2. ArticleLFM2.5-8B-A1B: An Even Better On-Device Mixture-of-Expertsliquid.ai · 2026-05-28 · official
  3. ArticleLFM2.5-2.6B: On-Device Agentsdocs.liquid.ai · 2026-08-04 · official
  4. ArticleLFM2.5-Encoders for Fast Long-Context Inference on CPUhuggingface.co · 2026-08-06 · 1 HN pts
  5. ArticleLFM2.5-2.6B: Deploy Agents Everywhereliquid.ai · 2026-08-04 · official
  6. ArticleAccomplishing More with LFM2.5-VL-DSpark2tinteractive.com · 2026-09-24
  7. ArticleLFM2.5-VL-450M: Structured Visual Intelligence, Edge to Cloudliquid.ai · 2026-04-27 · official
  8. DiscussionLFM2.5-350Mtwitter.com · 2026-04-02 · 3 HN pts

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.

No API provider serves it, so its downloads are people running it themselves: 87 quantized builds and 19 Spaces.

  • As a service0no API provider listed
  • On own hardware87quantized builds
  • Built on13483 fine-tunes · 50 adapters · 1 merges
  • In apps19Hugging Face Spaces

Top Spaces running it: webml-community/lfm2.5-webgpu-summarizer ♥ 33 · geocine/MiniMax-H3-Prompt-Enhancer ♥ 23 · stanley-00/slm-testing ♥ 9 · FlameF0X/lfm2 ♥ 7

Will it fit?

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

Fits a 16 GB card at INT4

About 0.2 GB at INT4, 0.4 GB at FP8, 0.9 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 0.2 GB of weights (estimate; KV cache and activations extra)INT40.2 GBFP8: about 0.4 GB of weights (estimate; KV cache and activations extra)FP80.4 GBBF16: about 0.9 GB of weights (estimate; KV cache and activations extra)BF160.9 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 0.2 GB of weights (estimate; KV cache and activations extra)INT40.2 GBFP8: about 0.4 GB of weights (estimate; KV cache and activations extra)FP80.4 GBBF16: about 0.9 GB of weights (estimate; KV cache and activations extra)BF160.9 GB

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 (2026-03-31), amber triangles are 4 dated articles (each links to it), grey ticks are repo commits.

AprMayJunJulAugSepOct2026-08-05: 1 repo commit2026-08-03: 1 repo commit2026-07-09: 1 repo commit2026-07-02: 5 repo commits2026-06-03: 1 repo commit2026-04-01: 4 repo commits2026-03-31: 4 repo commitsLFM2.5-DSpark: Up to 3.2x Faster Inference from H100 to MacB · 2026-08-21LFM2.5-Encoders for Fast Long-Context Inference on CPU · 2026-08-06LFM2.5-350M: No Size Left Behind · 2026-04-01LFM2.5-350M · 2026-04-02Weights 2026-03-31 · 2026-03-31Weights 2026-03-31now
AprMayJunJulAugSepOct2026-08-05: 1 repo commit2026-08-03: 1 repo commit2026-07-09: 1 repo commit2026-07-02: 5 repo commits2026-06-03: 1 repo commit2026-04-01: 4 repo commits2026-03-31: 4 repo commitsLFM2.5-DSpark: Up to 3.2x Faster Inference from H100 to MacB · 2026-08-21LFM2.5-Encoders for Fast Long-Context Inference on CPU · 2026-08-06LFM2.5-350M: No Size Left Behind · 2026-04-01LFM2.5-350M · 2026-04-02Weights 2026-03-31 · 2026-03-31Weights 2026-03-31now

What people built on it

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

221 derivatives on the Hub, led by 87 quantized

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

87 quantized (Hugging Face model tree)83 fine-tunes (Hugging Face model tree)1 merge (Hugging Face model tree)50 adapters (Hugging Face model tree)87 quantized83 fine-tunes1 merge50 adapters
87 quantized (Hugging Face model tree)83 fine-tunes (Hugging Face model tree)1 merge (Hugging Face model tree)50 adapters (Hugging Face model tree)87 quantized83 fine-tunes1 merge50 adapters
GGUF LiquidAI/LFM2.5-350M-GGUF for llama.cpp, Ollama, LM Studio 330.5k downloads MLX LiquidAI/LFM2.5-350M-MLX-8bit for Apple silicon 9.9k downloads

All variants →

Reference

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

Benchmarks

4 rows

Benchmarks 4 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (4)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.
BenchmarkValueSetting / notesReported byDate
author-reported ifstruct_v1 (LiquidAI/ifstruct-v1.0)44.9+ IFStruct RLLiquid AI — IFStruct v1.0 blog2026-06-30
author-reported ifstruct_v1 (LiquidAI/ifstruct-v1.0)21.1—Liquid AI — IFStruct v1.0 blog (LFM2.5-350M)2026-06-30
author-reported diamond (Idavidrein/gpqa)30.64—Model Card2026-03-31
author-reported mmlu_pro (TIGER-Lab/MMLU-Pro)20.01—Model Card2026-03-31

What the card says

2 cited sentences

Summary generated, cited

Designed for on-device deployment, this instruction-tuned model supports a context length of 32,768 tokens.[1][2] The model is released under the lfm1.0 license and is trained on a budget of 28T tokens.[3][4]

  1. “designed for **on-device deployment**”
  2. “Context length: 32,768 tokens”
  3. license
  4. “Training budget: 28T tokens”

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

Use-case evidence

4 use cases

What it is good at 4 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

Pipeline tagtext-generation
Librarytransformers
ArchitecturesLfm2ForCausalLM
Model typelfm2
Dtypebfloat16
Params by dtypeBF16: 354,483,968
Size bandS
MoEno
Context length128,000 (config.json max_position_embeddings)
Languagesen, ar, zh, fr, de, ja, ko, es, pt
PapersarXiv:2511.23404
Licenselfm1.0 (other)
Hub snapshot2026-10-25

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 LFM2.5-350M: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)0.9 GB · Edge (<4 GB)
FP8 (1 byte/param)0.4 GB · Edge (<4 GB)
INT4 (0.5 bytes/param)0.2 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/LFM2.5-350M-Base (same organization) Family: LiquidAI/LFM2.5 221 derivatives on the Hub

Base model: LiquidAI/LFM2.5-350M-Base

quantized: 87finetune: 83adapter: 50merge: 1

Top variants (40 shown, ranked by downloads)

Relation
Quant
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
LiquidAI/LFM2.5-350M-GGUF quantized GGUF — — other LiquidAI 108 330.5k
distil-labs/distil-lfm25-shellper finetune — 354M S other distil-labs 11 278.3k
OrdenWills/LFM2.5-350M-home-assistant-sft adapter GGUF 353M S apache-2.0 OrdenWills 5 17.1k
onnx-community/LFM2.5-350M-ONNX quantized — — — other onnx-community 5 16.6k
SupraLabs/Supra-Title-350M-exp-GGUF quantized GGUF — — apache-2.0 SupraLabs 50 10.3k
LiquidAI/LFM2.5-350M-MLX-8bit quantized MLX 354M S other LiquidAI 14 9.9k
LiquidAI/LFM2.5-350M-ONNX quantized — — — other LiquidAI 21 5.4k
squ11z1/claude-oss-350m finetune GGUF 354M S other squ11z1 11 5.7k
LiquidAI/LFM2.5-350M-MLX-4bit quantized MLX 354M S other LiquidAI 13 5.4k
mradermacher/LFM2.5-350M-GGUF quantized GGUF — — other mradermacher 1 5.4k
mradermacher/LFM2.5-350M-i1-GGUF quantized GGUF — — other mradermacher 1 5.1k
ZuzeTt/LFM2.5-VL-450M-GGUF quantized GGUF — — other ZuzeTt 1 4.1k
lmstudio-community/LFM2.5-350M-GGUF quantized GGUF — — other lmstudio-community 4 3.4k
LiquidAI/LFM2.5-350M-MLX-bf16 finetune MLX 354M S other LiquidAI 8 2.4k
lazos/lfm2.5-350m-frontend-agent finetune GGUF 354M S other lazos 2 2.3k
MihaiPopa-1/LFM2.5-350M-heretic-xagressive finetune — 354M S other MihaiPopa-1 0 2k
LiquidAI/LFM2.5-350M-MLX-6bit quantized MLX 354M S other LiquidAI 11 1.4k
Iam314rock/Qlora-lfm2.5-350m-mental-health-ONNX quantized — — — other Iam314rock 1 1.9k
OpenVINO/LFM2.5-350M-fp16-ov finetune -fp16 — — other OpenVINO 4 1.5k
LiquidAI/LFM2.5-350M-MLX-5bit quantized MLX 354M S other LiquidAI 9 1.2k
KoarAI/LFM2.5-350M-Thinking-0003 finetune — 353M S apache-2.0 KoarAI 1 1.6k
bitsydarel/road-freight-voice-assistant-350m quantized GGUF — — other bitsydarel 2 1.5k
itsZyn/ZynDwarf-1.1 quantized GGUF 354M S apache-2.0 itsZyn 1 1.4k
BenjaminHelle/LFM2.5-350M-coder quantized GGUF — — — BenjaminHelle 1 1.4k
Piecrust/Spike-350M-GGUF quantized GGUF — — other Piecrust 0 1.4k
YoussefElsafi/Aiko-350M-GGUF quantized GGUF — — apache-2.0 YoussefElsafi 1 1.3k
mlx-community/LFM2.5-350M-bf16 finetune -bf16 354M S other mlx-community 1 1.3k
prithivMLmods/LFM2.5-350M-F32-GGUF quantized GGUF — — apache-2.0 prithivMLmods 2 1.2k
DireDreadlord/Gallium-350M finetune — 353M S other DireDreadlord 11 733
MihaiPopa-1/LFM2.5-350M-heretic finetune — 354M S other MihaiPopa-1 3 1.1k
anudit/lfm25-strudel quantized — 354M S other anudit 0 1.3k
OpenVINO/LFM2.5-350M-int8-ov quantized int8 — — other OpenVINO 5 999
mlx-community/LFM2.5-350M-6bit quantized — 354M S other mlx-community 0 1.2k
DedeProGames/NTX-350m-Preview finetune — 354M S apache-2.0 DedeProGames 1 1.1k
itsZyn/ZynDwarf-1.0 quantized GGUF 354M S other itsZyn 2 1k
mlx-community/LFM2.5-350M-8bit quantized -8bit 354M S other mlx-community 0 1.1k
lewisdog/lfm2.5-350m-cogs-ask finetune — 354M S other lewisdog 0 1.1k
Null-Guard/LFM2.5-350M-Uncensored-GGUF quantized GGUF — — apache-2.0 Null-Guard 0 1.1k
Q1ngMang/LFM2.5-350M-PT-zh_CN finetune — 354M S cc-by-sa-3.0 Q1ngMang 1 1k
Indexnusrefather/Super-Slop-Machina-Macro-Roleplay-Instruct-350M-0812 finetune GGUF 354M S other Indexnusrefather 1 1k

Repository changelog

17 commits
changed in the last 7 daysOther: 2026-10-24, 2026-10-24all changes

Repository changelog 17 commits

DateCommitSHA
2026-08-05 Fix chat template: render assistant tool_calls (#12) 9e6c6cc
2026-08-03 Update README.md 36aa424
2026-07-09 Streamline special tokens (#10) b9d6e4e
2026-07-02 Delete .eval_results/ifstruct_v1.yaml 5b423cd
2026-07-02 Add IFStruct v1.0 evaluation result (#9) e8b29ac
2026-07-02 Add IFStruct v1.0 evaluation result (#8) 476054b
2026-07-02 Update .eval_results/ifstruct_v1.yaml 083022f
2026-07-02 Create ifstruct_v1.yaml 6e60feb
2026-06-03 Update README.md (#5) 7728373
2026-04-01 Add GPQA Diamond evaluation result (#2) 7081022
2026-04-01 Add MMLU-Pro evaluation result (#3) 1575d1b
2026-04-01 Update README.md b624815
2026-04-01 Update README.md 20f92cc
2026-03-31 Update README.md edba0e7
2026-03-31 Update README.md 302de0b
2026-03-31 Add files using upload-large-folder tool 55e5a53
2026-03-31 initial commit 0664b7f

Articles and coverage

10 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.

Official3

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-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.
LFM2.5-350M: No Size Left Behind
liquid.ai2026-04-01officialHN 3 pts · 1 comments1,629 words

Community1

LFM2.5-350M
twitter.com2026-04-02HN 3 pts · 0 comments
All coverage tracked for this model (10) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
Accomplishing More with LFM2.5-VL-DSpark2tinteractive.com2026-09-242026-10-07hn
LFM2.5-DSpark: Up to 3.2x Faster Inference from H100 to MacBliquid.ai official2026-08-212026-10-07hn
LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edgeliquid.ai official2026-08-122026-10-07hn
LFM2.5-Encoders for Fast Long-Context Inference on CPUhuggingface.co2026-08-062026-10-07hn
LFM2.5-2.6B: On-Device Agentsdocs.liquid.ai official2026-08-042026-10-07hn
LFM2.5-2.6B: Deploy Agents Everywhereliquid.ai official2026-08-042026-10-07hn
LFM2.5-8B-A1B: An Even Better On-Device Mixture-of-Expertsliquid.ai official2026-05-282026-10-07hn
LFM2.5-VL-450M: Structured Visual Intelligence, Edge to Cloudliquid.ai official2026-04-272026-10-07hn
LFM2.5-350Mtwitter.com2026-04-022026-10-07hn
LFM2.5-350M: No Size Left Behindliquid.ai official2026-04-012026-10-07hn

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).

Similar models

3 models

Similar models

  • SmolLM2-360M · 362M
    same size band (S), same task (text-generation), permissive license
  • gpt2-medium · 380M
    same size band (S), same task (text-generation), permissive license
  • Qwen2.5-0.5B-Instruct · 494M
    same size band (S), same task (text-generation), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:39 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:10 UTC · Last verified 2026-10-25

Follow LFM2.5-350M

Watch this model

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