Model Follower.

Up to date

Rebuilt
2026-11-15 07:14 UTC
Release radar
108 on watch
Hub snapshot
2026-11-15
Re-checked
daily · every number sourced and dated

Models › LFM3 family

LFM3-1.7B

Hugging Face ↗ Follow changes

LiquidAI lfm1.0 out 22 days ago

1.73B parameters, text in, text out.

Verified 2026-11-15 06:56 UTC Rank #200 in LLMs · score 88.3 method v1 Sources 3 Downloads 30d 100.6k · #305 of 618 tracked

Changed this week: 4 new quantized builds on the Hub (131 in all).2026-11-15 · 6 new quantized builds on the Hub (127 in all).2026-11-13 · 4 new quantized builds on the Hub (121 in all).2026-11-12 · 3 more

Can I run it?1.0GB at INT4Estimatefits a 16 GB cardmethod
Can I use it?lfm1.0Custom license; read the repo LICENSE before commercial hostinglicence file
Do people use it?100.6kdownloads, 30 days#305 of 618 we track · top 49%+249 likes, downloads +44% this week1,787 likesHugging Face

Worth reading

Picked by rule from 11 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Latest · 17 days ago

LFM3 Proves Small Models Can Hold Long Context, Not That They Can Use It

LFM3 is a collection of six open-weight models built on a 1.73B-parameter backbone, released on October 27, 2026.

  • Adapters written for standard transformers miss most of LFM3 because target modules skip the convolution blocks entirely.
snackonai.com 2026-10-29 3.1k words about the LFM3 family family

11 best of 11 tracked

The 3 picks above, plus these 8. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. CodeShow HN: LFM3 Video Chatgithub.com · 2026-10-28 · 5 HN pts
  2. ArticleBuild specialized multimodal models with LFM3huggingface.co · 2026-10-27 · shared by 2
  3. ArticleLFM3: Technical Deep Dive into Liquid AI's Tri-Hybrid Multimodal Family - Local AI Zonelocal-ai-zone.github.io · 2026-10-28
  4. ArticleLiquid AI's new LFM3 models put browser agents, voice ...venturebeat.com · 2026-10-28
  5. ArticleLFM3-1.7B-Base: What the Repo Says, What Is Unprovenorcarouter.ai · 2026-10-27
  6. ArticleLFM3 keeps 128K tokens of context in 1.6 GB, and its benchmark lead depends on a score for abstaining — sysnotesysnote.net · 2026-10-27
  7. ArticleLiquidAI/LFM3 · FastFlowLMfastflowlm.com · undated
  8. ArticleInside LFM3: Liquid AI's Latest Models Come to Oscillaoscilla.ai · undated

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: 152 quantized builds.

  • As a service0no API provider listed
  • On own hardware152quantized builds
  • Built on9567 fine-tunes · 27 adapters · 1 merges
  • In apps0Hugging Face Spaces

Will it fit?

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

Fits a 16 GB card at INT4

About 1.0 GB at INT4, 2.1 GB at FP8, 4.1 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.0 GB of weights (estimate; KV cache and activations extra)INT41.0 GBFP8: about 2.1 GB of weights (estimate; KV cache and activations extra)FP82.1 GBBF16: about 4.1 GB of weights (estimate; KV cache and activations extra)BF164.1 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.0 GB of weights (estimate; KV cache and activations extra)INT41.0 GBFP8: about 2.1 GB of weights (estimate; KV cache and activations extra)FP82.1 GBBF16: about 4.1 GB of weights (estimate; KV cache and activations extra)BF164.1 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.

Coverage peaked at launch and is still arriving

Teal is the weights (2026-10-23), amber triangles are 8 dated articles (each links to it).

NovIntroducing LFM3: Every Modality for Specialized Intelligence · 2026-10-27Show HN: LFM3 Video Chat · 2026-10-28Build specialized multimodal models with LFM3 · 2026-10-27LFM3 Proves Small Models Can Hold Long Context, Not That They Can Use It · 2026-10-29LFM3: Technical Deep Dive into Liquid AI's Tri-Hybrid Multimodal Family - Local AI Zone · 2026-10-28Liquid AI's new LFM3 models put browser agents, voice ... · 2026-10-28LFM3-1.7B-Base: What the Repo Says, What Is Unproven · 2026-10-27LFM3 keeps 128K tokens of context in 1.6 GB, and its benchmark lead depends on a score for abstaining — sysnote · 2026-10-27Weights 2026-10-23 · 2026-10-23Weights 2026-10-23now
NovIntroducing LFM3: Every Modality for Specialized Intelligence · 2026-10-27Show HN: LFM3 Video Chat · 2026-10-28Build specialized multimodal models with LFM3 · 2026-10-27LFM3 Proves Small Models Can Hold Long Context, Not That They Can Use It · 2026-10-29LFM3: Technical Deep Dive into Liquid AI's Tri-Hybrid Multimodal Family - Local AI Zone · 2026-10-28Liquid AI's new LFM3 models put browser agents, voice ... · 2026-10-28LFM3-1.7B-Base: What the Repo Says, What Is Unproven · 2026-10-27LFM3 keeps 128K tokens of context in 1.6 GB, and its benchmark lead depends on a score for abstaining — sysnote · 2026-10-27Weights 2026-10-23 · 2026-10-23Weights 2026-10-23now

What people built on it

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

247 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.

152 quantized (Hugging Face model tree)67 fine-tunes (Hugging Face model tree)1 merge (Hugging Face model tree)27 adapters (Hugging Face model tree)152 quantized67 fine-tunes1 merge27 adapters
152 quantized (Hugging Face model tree)67 fine-tunes (Hugging Face model tree)1 merge (Hugging Face model tree)27 adapters (Hugging Face model tree)152 quantized67 fine-tunes1 merge27 adapters
GGUF LiquidAI/LFM3-1.7B-GGUF for llama.cpp, Ollama, LM Studio 342.9k downloads MLX LiquidAI/LFM3-1.7B-MLX-4bit for Apple silicon 15.8k downloads NVFP4 LiquidAI/LFM3-1.7B-NVFP4 for vLLM, SGLang 5k downloads

All variants →

Reference

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

Use-case evidence

3 use cases

What it is good at 3 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

14 fields

Listed in LLMs

Facts

Pipeline tagtext-generation
Librarytransformers
ArchitecturesLfm3ForCausalLM
Model typelfm3
Params by dtypeBF16: 1,728,801,536
Size bandS
MoEno
Licenselfm1.0 (other)
Hub snapshot2026-11-15

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for LFM3-1.7B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)4.1 GB · Single GPU (<24 GB)
FP8 (1 byte/param)2.1 GB · Edge (<4 GB)
INT4 (0.5 bytes/param)1.0 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/LFM3-1.7B-Base (same organization) Family: LiquidAI/LFM3 247 derivatives on the Hub

Base model: LiquidAI/LFM3-1.7B-Base

quantized: 152finetune: 67adapter: 27merge: 1

Top variants (40 shown, ranked by downloads)

Relation
Quant
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
LiquidAI/LFM3-1.7B-GGUF quantized GGUF — — other LiquidAI 293 342.9k
AtomicChat/LFM3-1.7B-Uncensored-GGUF quantized GGUF — — other AtomicChat 320 151.4k
turintech/LFM3-1.7B-Artemis-GGUF quantized GGUF — — other turintech 10 22.3k
LiquidAI/LFM3-1.7B-MLX-4bit quantized MLX 1.73B S other LiquidAI 20 15.8k
AtomicChat/LFM3-1.7B-GGUF quantized GGUF — — other AtomicChat 26 10.2k
bartowski/LFM3-1.7B-GGUF quantized GGUF — — other bartowski 6 9.8k
lmstudio-community/LFM3-1.7B-GGUF quantized GGUF — — — lmstudio-community 4 8.1k
LiquidAI/LFM3-1.7B-MLX-8bit quantized MLX 1.73B S other LiquidAI 14 7.1k
LiquidAI/LFM3-1.7B-MLX-bf16 quantized MLX 1.73B S other LiquidAI 16 6k
LiquidAI/LFM3-1.7B-NVFP4 quantized NVFP4 1.73B S other LiquidAI 27 5k
NANI-Nithin/LFM3-1.7B-GGUF quantized GGUF — — — NANI-Nithin 0 5.9k
Infatoshi/phonon-refiner-1.7b finetune GGUF 1.73B S other Infatoshi 14 4.6k
software-mansion/LFM3-ExecuTorch quantized — — — other software-mansion 1 5.2k
saidutta69/LFM3-1.7B-heretic quantized GGUF 1.73B S other saidutta69 3 4.9k
mohommed/LFM3-1.7B-Frontier-SmartAgent quantized GGUF 1.73B S apache-2.0 mohommed 14 3.9k
AJ-Gonzalez/coder-lfm3-1.7b adapter GGUF 1.73B S other AJ-Gonzalez 18 3.6k
Soulfate24/LFM3-1.7B-Paretrix-GGUF quantized GGUF — — other Soulfate24 1 4.3k
LiquidAI/LFM3-1.7B-MLX-6bit quantized MLX 1.73B S other LiquidAI 10 3.8k
saidutta69/LFM3-1.7B-Claude-Fable5.1 adapter GGUF 1.73B S other saidutta69 6 3.8k
cybertruck32489/LFM3-1.7B-RD finetune — 1.70B S other cybertruck32489 3 3.8k
LiquidAI/LFM3-1.7B-FP8 quantized -FP8 1.73B S other LiquidAI 13 3.2k
mondk/LFM3-1.7B-Claude-Code-GGUF quantized GGUF — — other mondk 8 3.4k
Schnuckade/LFM-3-Coder-Claude-Distill quantized GGUF 1.73B S apache-2.0 Schnuckade 9 3.2k
litert-community/LFM3-1.7B-LiteRT quantized — — — other litert-community 8 3.2k
LiquidAI/LFM3-1.7B-ONNX quantized — — — other LiquidAI 15 2.8k
Arki05/LFM3-1.7B-GGUF-shards quantized GGUF — — other Arki05 2 3.3k
mkieffer/LFM3-FHIR-demo-v1 quantized GGUF — — other mkieffer 0 3.4k
orcarouter/LFM3-1.7B-abliterated finetune — 1.73B S other orcarouter 51 745
ngquocvinh/LFM3-1.7B-GGUF quantized GGUF — — other ngquocvinh 1 2.9k
rSalone/LFM3-1.7B-uncensored-abl48 finetune GGUF 1.73B S other rSalone 1 2.8k
ReStranger/LFM3-1.7B-Claude-Code-Abliterated-GGUF quantized GGUF — — other ReStranger 7 2.3k
AJ-Gonzalez/coder-lfm3-1.7b-k3.1-pro adapter GGUF 1.73B S other AJ-Gonzalez 25 1.3k
LiquidAI/LFM3-1.7B-MLX-5bit quantized MLX 1.73B S other LiquidAI 8 1.9k
mondk/LFM3-1.7B-Claude-Code-Safetensors finetune — 1.73B S other mondk 7 1.9k
llama-swap/LFM3-1.7B-HelpAgent finetune GGUF 1.73B S other llama-swap 2 2k
onnx-community/LFM3-1.7B-ONNX quantized — — — other onnx-community 5 1.8k
kaitchup/LFM3-1.7B-SLQ-GGUF quantized GGUF — — other kaitchup 13 1.4k
Arki05/LFM3-1.7B-GGUF quantized GGUF — — other Arki05 3 1.9k
OpenIntelligenceNet/LFM3-1.7B-Uncensored finetune — 1.73B S — OpenIntelligenceNet 10 1.5k
bowtiedrahman/LFM3-1.7B-abliterated quantized GGUF 1.73B S other bowtiedrahman 1 1.8k

Repository changelog

7 changes
changed in the last 7 daysRepo updated: 2026-11-11Other: 2026-11-15, 2026-11-13, 2026-11-12 +3all changes

Articles and coverage

11 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

Introducing LFM3: Every Modality for Specialized Intelligence
liquid.ai2026-10-27officialHN front pageabout the LFM3 familyHN 171 pts · 67 comments
Build specialized multimodal models with LFM3
huggingface.co2026-10-27authorityabout the LFM3 family2,579 words
LiquidAI launches LFM3, a specialized multimodal model family sharing a 2B parameter backbone across text, vision, audio, omni, and decision variants. The 1.7B base model features a 24-layer tri-hybrid architecture with a 256K context window, fitting in about 1 GB at 4-bit precision while using 35% fewer parameters than LFM2.5-2.6B. New in-house encoders include a 164M-parameter vision model requiring 2.2x fewer FLOPs than SigLIP2, and an audio encoder enabling time-to-first-token as low as 50 ms.written, reviewed, cited
  • The vision encoder uses 3x less compute than the SigLIP2 encoder in previous LFM VL models.
  • LiquidCodec turns generated tokens into speech at about 87x real time on a single CPU thread on Apple M4 Max.
  • The audio encoder is fully causal, so the model can prefill while you speak.
LFM3-1.7B - Liquid AI Docs
docs.liquid.aiofficial183 words

Analysis7

Show HN: LFM3 Video Chat
github.com2026-10-28authorityabout the LFM3 familyHN 5 pts · 1 comments830 words
LFM3-Omni-2B enables video chat with text and audio replies, supporting interruption via speech or button press. The project offers ONNX and Apple Silicon MLX backends, with Smart Turn v3.2 running on the CPU to distinguish thinking pauses from sentence ends. Localhost access requires no SSL, but remote connections need a certificate valid for 365 days generated via openssl.written, reviewed, cited
  • The ONNX backend runs without PyTorch by using q4 quantization for CPU or q4f16 for NVIDIA GPUs.
  • The experimental Apple Silicon version utilizes the 8-bit LFM3-Omni model with the MLX framework.
  • The Smart Turn model folder must be preserved to reuse downloads, as deleting it removes the local copy.
LFM3 Proves Small Models Can Hold Long Context, Not That They Can Use It
snackonai.com2026-10-29about the LFM3 family3,150 words
LFM3 is a collection of six open-weight models built on a 1.73B-parameter backbone, released on October 27, 2026. The architecture uses mostly convolution and only six global attention layers to achieve a small memory footprint while supporting up to 256K context, though the hosted API lists 128K. Thinking is on by default, and the card warns that enable_thinking=False can behave inconsistently. The models were pre-trained on 22T tokens and are licensed under the LFM Open License v1.0.written, reviewed, cited
  • Adapters written for standard transformers miss most of LFM3 because target modules skip the convolution blocks entirely.
LFM3: Technical Deep Dive into Liquid AI's Tri-Hybrid Multimodal Family - Local AI Zone
local-ai-zone.github.io2026-10-28about the LFM3 family9,034 words
Liquid AI released LFM3 on October 27, 2026, a tri-hybrid multimodal family built on a single 1.7B backbone trained on 22T tokens. The architecture uses 13 gated short convolutions, 5 sliding-window attention layers, and 6 global attention layers to achieve faster edge inference.written, reviewed, cited
  • The convolution kernel size increased to 7, up from 3 in the previous LFM2 generation.
  • A full 128K-token context requires 1.6 GB of KV cache, significantly less than Qwen3.5-4B.
  • Inference relies on dedicated width-7 Triton kernels because standard causal_conv1d kernels only accept widths 2–4.
Liquid AI's new LFM3 models put browser agents, voice ...
venturebeat.com2026-10-28about the LFM3 family
LFM3-1.7B-Base: What the Repo Says, What Is Unproven
orcarouter.ai2026-10-272,062 words
LFM3 keeps 128K tokens of context in 1.6 GB, and its benchmark lead depends on a score for abstaining — sysnote
sysnote.net2026-10-27about the LFM3 family1,009 words
LiquidAI/LFM3 · FastFlowLM
fastflowlm.comabout the LFM3 family319 words
All coverage tracked for this model (11) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
LFM3-1.7B - Liquid AI Docsdocs.liquid.ai official—2026-11-01exa
LiquidAI/LFM3 · FastFlowLMfastflowlm.com—2026-11-01exa
Inside LFM3: Liquid AI's Latest Models Come to Oscillaoscilla.ai—2026-11-01exa
LFM3 Proves Small Models Can Hold Long Context, Not That They Can Use Itsnackonai.com2026-10-292026-11-01exa
Show HN: LFM3 Video Chatgithub.com2026-10-282026-10-30hn
LFM3: Technical Deep Dive into Liquid AI's Tri-Hybrid Multimodal Family - Local AI Zonelocal-ai-zone.github.io2026-10-282026-11-01exa
Liquid AI's new LFM3 models put browser agents, voice ...venturebeat.com2026-10-282026-11-01exa
LFM3-1.7B-Base: What the Repo Says, What Is Unprovenorcarouter.ai2026-10-272026-11-01exa
Introducing LFM3: Every Modality for Specialized Intelligenceliquid.ai official2026-10-272026-10-30hn
Build specialized multimodal models with LFM3huggingface.co2026-10-272026-11-01exa
LFM3 keeps 128K tokens of context in 1.6 GB, and its benchmark lead depends on a score for abstaining — sysnotesysnote.net2026-10-272026-11-01exa

Similar models

3 models

Similar models

  • DeepSeek-R1-Distill-Qwen-1.5B · 1.78B
    same size band (S), same task (text-generation), more providers, permissive license
  • Qwen3-1.7B-Base · 1.72B
    same size band (S), same task (text-generation), more providers, permissive license
  • Qwen2.5-1.5B-Instruct · 1.54B
    same size band (S), same task (text-generation), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-11-15 · Providers: 2026-11-15 06:55 UTC · Articles: 2026-11-15 06:56 UTC · Summary: 2026-10-30 07:09 UTC · Last verified 2026-11-15 06:56 UTC

Follow LFM3-1.7B

Watch this model

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