Skip to content
FrontierDot

Liquid LFM3 Open Weights Timeline: Six ~2B Models From One Tri-Hybrid Backbone (Oct 27, 2026)

Live timeline · last updated · 2 entries in the update log

On October 27 Liquid published six LFM3 models around the 2B range, weights on Hugging Face the same day. Read it as an edge executor family. Tool rows are Liquid-reported. The license name is LFM Open License v1.0.

Timeline

Liquid AI released LFM3 on Oct 27, 2026 as six open-weight ~2B checkpoints from LFM3-1.7B-Base under LFM Open License v1.0. The LICENSE file defines a commercial-use Threshold of $10,000,000 in annual revenue. Tool-use scores on the card are Liquid-reported, and the card says the 1.7B text model is not recommended for agentic coding or complex multi-step reasoning.

Read on 2026-10-29 (UTC+8): Liquid’s Oct 27 blog at liquid.ai/blog/lfm3 (including the 1.6 GB KV-cache sentence), the Hugging Face blog “Build specialized multimodal models with LFM3” (Published October 27, 2026), the LiquidAI/LFM3-1.7B model card, its LICENSE file (LFM Open License v1.0), Liquid’s license guide (last updated October 15, 2026), the Hacker News thread item?id=50262229, llama.cpp PR #31252, Ollama v0.40.12 release notes, and @liquidai’s Introducing LFM3 post (created 2026-10-27T17:07:45.000Z). X times are UTC as X reports them. Architecture claims and benchmark scores are Liquid-reported. We did not run the models.

Timeline

  1. Liquid AI publishes “Introducing LFM3: Every Modality for Specialized Intelligence,” dated OCT 27, 2026. The post describes a multimodal family in the 2B parameter range branched from LFM3-1.7B-Base, a new tri-hybrid architecture, 22T tokens of pre-training, a 256K context window for the text model, about 1 GB at 4-bit for the text checkpoint, and six models: LFM3-1.7B, LFM3-VL-2B, LFM3-Audio-2B, LFM3-Omni-2B, d1-2B and d1-Omni-2B. It says the base and post-trained models are open-weight on Hugging Face, Liquid’s Playground and the Liquid AI API.

    Liquid AI blog: Introducing LFM3

  2. llama.cpp pull request #31252, adding LiquidAI/LFM3-1.7B and LiquidAI/LFM3-VL-2B, is merged. GitHub records the merge at 2026-10-27T16:06:41Z. The pull request says video input support will be added later. We did not build llama.cpp on Oct 29.

    llama.cpp PR #31252 (LFM3 text and VL)

  3. At 17:07:45 UTC the official @liquidai account posts “Introducing LFM3…” on X, matching the same-day blog announcement.

    Liquid AI on X: Introducing LFM3 (Oct 27)

  4. Hugging Face publishes “Build specialized multimodal models with LFM3,” marked Published October 27, 2026. The post repeats the six-model family, the tri-hybrid layer mix, the 22T pretrain and 256K context figures, and links the LFM3 collection and per-model cards.

    Hugging Face blog: Build specialized multimodal models with LFM3

  5. Ollama’s v0.40.12 release notes list LFM3 text, VL, audio and omni, and say lfm3 and d1 2b are supported on Apple hardware in that release, with Linux and Windows support described as coming soon. We read those notes and the library page on Oct 29. We did not pull the model, and we did not read every later pre-release.

    Ollama v0.40.12 release notes

  6. When we checked the LiquidAI/LFM3-1.7B repository API, lastModified was approximately 2026-10-28T17:16:37Z. The public announcement date stays Oct 27. Treat Oct 28 as a card or weight refresh timestamp.

    Hugging Face: LiquidAI/LFM3-1.7B model card

What Liquid shipped on October 27

The Liquid blog post frames LFM3 as the next Liquid Foundation Model generation for specialized intelligence: multimodal by design, with text, vision, audio, omni and decision models on one shared backbone. The first release covers models in the 2B parameter range. Liquid says the models are open-weight and available the same day on Hugging Face, its Playground and the Liquid AI API, and points readers to docs for local run and fine-tune steps.

The six named checkpoints are LFM3-1.7B (text), LFM3-VL-2B (vision, including video input for the first time in an LFM vision model), LFM3-Audio-2B, LFM3-Omni-2B (Liquid’s first omni model), plus decision models d1-2B and d1-Omni-2B. The Hugging Face blog table writes the last of those as d1-omni-2B; both spellings refer to the audio-capable decision model built on LFM3-Omni-2B. Liquid reports that LFM3-1.7B scores above LFM2.5-2.6B on Liquid’s own highlight comparison while using 35% fewer parameters, that the text model fits in about 1 GB at 4-bit, and that every model can be fine-tuned on a single GPU. Those are vendor claims. We did not reproduce them.

Day-zero partners named on the Liquid post include Ambarella, AMD, Arm, Intel via OpenVINO, NVIDIA, OpenRouter, Qualcomm via Qualcomm AI Hub and GenieX, and Vercel, with Mercedes-Benz and Shopify mentioned for pre-release testing. Builder-ecosystem names on the same page include Hugging Face, llama.cpp, LM Studio, MLX, Ollama, vLLM and others. Front lists those as Liquid’s partner line. A dated runtime page is what settles support for a given stack.

Architecture, as Liquid describes it

Liquid describes a tri-hybrid stack for LFM3-1.7B-Base: 24 layers made of 13 gated short convolutions (kernel size 7, up from 3 in LFM2), 5 sliding-window attention layers with a 256-token window, and 6 global attention layers. The Hugging Face card lists 1.73B parameters for the post-trained text model, a 256K-token context (input plus output), and a 128,000-token vocabulary shared with LFM2.5-8B-A1B. Extra design notes on the blog include head-wise gated attention, partial RoPE on global layers, and sandwich norm.

Pre-training is stated as 22T tokens, with a final 3.5T mid-training stage that raises the context window to 256K and increases the share of agentic traces and tool calls. Post-training is described in four stages: supervised fine-tuning, mixed reinforcement learning, multi-domain on-policy distillation, and agentic RL inside real harnesses. For anyone sizing a phone or laptop deploy, the actionable figures Liquid emphasizes are the ~1 GB 4-bit footprint and the 256K window; for anyone reading the score tables, the actionable habit is to treat every row as Liquid-reported until a third party reruns the harness.

The same Liquid blog says a full 128K-token context needs 1.6 GB of KV cache for the text model, and in that sentence names 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B. The 256K figure is the published context length. The about-1 GB figure is the 4-bit weight pitch. The cache is extra, and a 256K window scales past the 128K cache number Liquid printed. We did not measure KV cache. Vision and audio get their own encoder notes on the launch pages: a Liquid vision encoder described as using less compute than the prior SigLIP2 path, and an audio path with a causal encoder plus LiquidCodec for speech.

License: LFM Open License v1.0

The LFM3-1.7B card names the license LFM Open License v1.0. We read that LICENSE file on Oct 29, 2026. Section 1 defines “Threshold” as annual revenue of 10 million United States dollars ($10,000,000) or more. Section 5 says the commercial-use rights under this license are conditioned on not exceeding that Threshold, and that commercial use by a legal entity that exceeds it is not licensed under the agreement. Qualified non-profits, as the file defines them, are addressed in section 5(c) for non-commercial or research purposes. Apache 2.0 and MIT are different license texts.

Liquid’s LFM Open License user guide, marked last updated October 15, 2026, states the same threshold in Liquid’s plain language: free commercial use under $10 million USD annual revenue, and a separate commercial license at $10 million USD or more, via sales@liquid.ai. The guide is Liquid’s explanation of Liquid’s terms. It is not FrontierDot legal advice. Read the LICENSE file and the guide before you ship a product, fine-tune for a paid service, or host the weights.

“Open-weight” on Liquid’s pages means the checkpoints are downloadable and the company says you can fine-tune and deploy them. The commercial threshold lives in the LICENSE file and in that guide. Hugging Face availability and the LFM Open License v1.0 name should be read together.

Scores, as Liquid reports them

Liquid’s blog and the LFM3-1.7B card publish a wide agent, tool-use, coding, instruction-following, knowledge and long-context table against LFM2.5-2.6B, MiniCPM5 2B, Gemma 4 E4B, Qwen3.5 4B and Qwen3.5 9B. Examples the card leads with include LiveClaw 45.60, WebArena-Verified 38.42, BFCL V4 71.39, ToolSandbox 80.36, SWE-bench Verified (OpenCode) 43.40 and an unweighted Average of 49.37. The same Average row lists Qwen3.5 9B at 47.24. Vision, audio, omni and decision scores appear on their respective cards and on the launch blog. The card says the text model is not recommended for agentic coding or complex multi-step reasoning because of its small size, and it says to use retrieval or search for knowledge-heavy tasks.

Counting the scored rows above that Average on the card we read, the LFM3-1.7B column is higher than Qwen3.5 9B on 14 rows and lower on 15. Tool rows such as BFCL V4 (71.39 versus 62.83) and ToolSandbox (80.36 versus 76.41) sit in the higher set. Knowledge and long-context rows such as GPQA Diamond (63.37 versus 80.40) and RULER 128K (65.56 versus 87.11) sit in the lower set. The Average is Liquid’s arithmetic. The 14 and 15 are our count of those printed cells. We did not rerun the harness.

Serving notes on the card and blog describe GGUF, MLX, ONNX, BF16, FP8 and NVFP4 paths and give Liquid’s own edge and GPU throughput claims, including the about-1 GB 4-bit phone-or-laptop pitch. Treat throughput numbers as Liquid-reported until you measure the checkpoint on your own stack. The reason to open this card is the tool and schema rows. The average hides the knowledge and long-context rows where Liquid’s own table is behind the 9B column.

Runtime reality on 2026-10-29

Partner logos are not a runtime matrix. On Oct 29 we read the public pages below and did not install the stacks, pull weights, or time tokens per second. A later pre-release can move a row. This table is a reading note from that hour.

RuntimeWhat the public page saidWhat we did not do
llama.cppPR #31252 merged 2026-10-27T16:06:41Z for LFM3-1.7B and LFM3-VL-2B, with video input still later. b11941 notes add LFM3-Audio-2B and LFM3-Omni-2B in #31257.Build or run llama.cpp on 2026-10-29.
Ollamav0.40.12 notes say Apple hardware for lfm3 and d1 2b, with Linux and Windows coming soon. The library page lists lfm3.Pull the model, or read every pre-release after v0.40.12.
LM StudioThe LFM3-1.7B card names LM Studio next to GGUF, with llama.cpp and Ollama.Open LM Studio’s catalog on 2026-10-29.

What launch-week videos claim

Liquid’s own upload is lab marketing: Introducing LFM3. Creator titles from the same week include Ray Codes, “The 1GB Local AI Agent That Beats Qwen 9B”, and CodeMotion, “This 1GB AI Beat a 9B on Its Own Scoreboard”.

Those creator titles are creator claims. The scoreboard they point at is Liquid’s table. The LFM3-1.7B card also says the model is not recommended for agentic coding or complex multi-step reasoning. The license threshold and the modality pick are absent from the titles. We did not treat the videos as a benchmark, and we did not re-count views.

What we could not verify

  • Independent benchmark scores for any LFM3 model. Every number on the blog and cards is Liquid-reported. The 14-row and 15-row split is our count of Liquid’s printed cells.
  • Live tokens per second on our machines. We did not run LFM3 on Oct 29, 2026.
  • Whether an Ollama pre-release after v0.40.12 dropped the Apple-hardware sentence. We read the v0.40.12 notes and the library page.
  • Which LFM3 checkpoints LM Studio listed on Oct 29, 2026. We did not open that catalog.
  • Why the LFM3-1.7B repository showed a lastModified stamp on Oct 28 after an Oct 27 announcement. We date the public release to Oct 27 from the blog, Hugging Face blog and X post.
  • Star counts, download rankings, third-party leaderboard positions, and YouTube view counts. None of those are measurements we re-counted.

Our take

LFM3 is a same-day edge executor family. The reason to care is the tool and schema rows on Liquid’s own table. Knowledge and long-context retrieval are where that table loses to Qwen3.5 9B: GPQA Diamond 63.37 versus 80.40, RULER 128K 65.56 versus 87.11, inside a 14-higher and 15-lower count of Liquid’s printed rows. The Hugging Face card says the 1.7B model is not recommended for agentic coding or complex multi-step reasoning because of its size. Keep the Oct 27 announcement as the launch day. The Oct 28 Hugging Face lastModified stamp is a refresh time, and a second launch would need its own primary URL.

Pick the modality checkpoint. On the Hacker News thread, Liquid’s samstevens wrote that the models are still stronger without extra modalities: LFM3-1.7B for pure text, LFM3-VL-2B over Omni for text and vision, and the Audio model over Omni for ASR and TTS. That comment is a vendor-staff statement on a public thread, not a third-party leaderboard. Defaulting to Omni throws away the specialist the same thread says still wins its lane.

Related reading on Front: the Reflection Beam timeline, the Meta Muse Spark Open timeline, the MiniMax M3.1 Flash timeline and the DeepSeek V4.1 Flash 1017 timeline track the other major open-weight drops of the month. Every post is on the Front blog index.

Keep reading on Front

Every post, newest first, is on the Front blog index.

FAQ

When did Liquid release LFM3?

October 27, 2026, per Liquid’s blog, the Hugging Face launch post dated the same day, and @liquidai’s Introducing LFM3 post at 17:07:45 UTC.

How many LFM3 models shipped in the first release?

Six around the 2B range, all branched from LFM3-1.7B-Base: LFM3-1.7B, LFM3-VL-2B, LFM3-Audio-2B, LFM3-Omni-2B, d1-2B and d1-Omni-2B, according to Liquid’s Oct 27 blog.

What is the LFM3 architecture?

Liquid describes a tri-hybrid of 13 gated short convolutions, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers (24 layers total), pretrained on 22T tokens, with a 256K text context.

Is LFM3 open source?

The weights are public on Hugging Face under the LFM Open License v1.0. Read the LICENSE file. Apache 2.0 and MIT are different license texts.

Should I use LFM3-Omni-2B?

Pick the modality checkpoint. On Hacker News item 50262229, Liquid’s samstevens wrote that LFM3-1.7B is stronger for pure text, LFM3-VL-2B is stronger than Omni for text and vision, and the Audio model is stronger than Omni for ASR and TTS. That is a vendor-staff comment, not our measurement.

Is LFM3 free for commercial use?

Read the LICENSE file and Liquid’s user guide, last updated October 15, 2026. The LICENSE defines Threshold as annual revenue of $10,000,000 or more. The guide says commercial use is free under $10 million USD annual revenue, and that $10 million USD or more needs a separate commercial license. That is Liquid’s statement of its terms, not FrontierDot legal advice.

Update log

  • — Deepened the same day: LFM Open License v1.0 as written in the LICENSE file and Liquid’s guide (threshold $10,000,000), the modality note on Hacker News, Liquid’s 1.6 GB KV-cache figure at 128K, and runtime pages we did not re-benchmark.
  • — First published. Checked Liquid’s LFM3 blog, Hugging Face blog and LFM3-1.7B card, Liquid’s news index, and @liquidai on X through Oct 29, 2026 (UTC+8).

Sources