Liquid AI LFM3-1.7B Explained: Six Models Sharing One 1.7B Backbone for On-Device Agents

Quick Answer

Liquid AI's six-model LFM3 family shares one 1.7B backbone: 256K context, about 1GB at 4-bit, built for on-device agents.

On October 27, 2026, Liquid AI released LFM3, and its logic runs against the month's bigger-is-better grain: instead of a larger frontier model, six models — text, vision, audio, omni, and decision — all built on one 1.7B backbone (LFM3-1.7B-Base, 1.73B parameters), aimed at agents that run on phones, laptops, and edge devices with data that never leaves the device. The base reached Hugging Face on October 26, weights are open, and official Playground and API access exist alongside. Caveats up front: every benchmark figure here is vendor-reported (preliminary) with no independent reproduction yet; and the license is not unconditionally free — above $10M in annual revenue it needs a separate arrangement.\n\nDisclaimer: this site is not affiliated with Liquid AI. This article is product-technology reporting and analysis, not procurement, deployment, or investment advice. Scores are vendor-reported and marked preliminary; license and pricing details follow official pages and may change.

1. What it is: six models, one 1.7B backbone, released two days ago

Get the release structure right first: this is not a single model, it is a family. The LFM3 release of October 27, 2026 contains six open-weight models, all branched from LFM3-1.7B-Base — the base itself reached Hugging Face on October 26, pre-trained on 22T tokens, up from about 15T for LFM2 — followed by a 3.5T-token mid-training stage that shifted the data mix toward agentic traces and tool calls while stretching context to 256K. The positioning is explicit: not a shrunken cloud flagship, but a starting point designed for the device where the application runs, with every model fine-tunable on a single GPU.\n\nOne line of company background: Liquid AI is an MIT CSAIL spin-off that raised $250M in a round led by AMD in late 2024 and previously open-sourced LFM2; launch partners include Mercedes-Benz (in-car voice assistant, targeting first production in North America from the second half of 2026) and Shopify for pre-release testing, with day-one partners Ambarella, AMD, Arm, Intel (OpenVINO), NVIDIA, OpenRouter, Qualcomm, and Vercel. The term is two days old, so there is no staleness question; one naming trap to avoid: only the text model is 1.7B, the rest are 2B class, so the family's “2B class” label does not mean six 1.7B models.

2. The tri-hybrid architecture: convolution plus sliding-window plus global attention, and why it can stay small

LFM3's architecture is tri-hybrid: across 24 blocks, 13 layers are gated short convolutions, 5 are sliding-window attention layers with a 256-token window, and 6 are global attention layers, with a kernel size of 7 and a 128K vocabulary. The intent is plain: convolutions cheaply handle local patterns, sliding windows cover medium-range dependencies, and only 6 global layers handle information that truly needs the whole sequence — cutting the most expensive attention to a quarter of the stack is how a 1.7B model carries a 256K context. The text model covers 16 languages; one coverage detail worth checking: per a European review, only 7 of those are EU official languages, so check your language list before building for European markets.\n\nFootprint and formats are the most practical part of this release: the text model is about 1GB at 4-bit (3.46GB as a BF16 file), with BF16, FP8, and NVFP4 builds for vLLM and SGLang, GGUF for llama.cpp and Ollama, MLX for Apple Silicon, and ONNX — every mainstream edge runtime can pick it up from day one. Liquid says the architecture search covered 14 device and backend combinations, including the Samsung Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI, and Qualcomm Hexagon NPU — “built for devices” was tuned against a device matrix before launch, not asserted after it.

3. Benchmarks: a big agentic jump, but entirely vendor-reported

The headline is WebArena-Verified: LFM3-1.7B scores 38.42 against 3.02 for LFM2.5-2.6B, about a 12.7-fold improvement on the vendor's figures (recomputed here: 38.42 / 3.02 ≈ 12.7). Other vendor-reported scores: 71.39 on BFCL V4 tool calling, 96.66 on IFEval, 80.36 on ToolSandbox, 52.83 on AA-LCR, and 45.60 on LiveClaw; parameters are about 35% fewer than the previous 2.6B generation (about 34.6% on nominal values). Speed figures come from the same vendor source: 1.8 to 2.8 times faster decoding than the next model across three on-device setups, and 3.6B tokens per day from a single NVIDIA B300.\n\nAll of this must stay labeled: Liquid AI's own release materials, no independent reproduction; and the 12.7-fold figure carries a low-base effect — a 3.02 starting point means the previous generation essentially lacked agentic ability, so the multiple flatters while the absolute 38.42 is the number to compare against peers. Until independent evaluations appear, these scores prove that the small-model agent path works, not that a 1.7B model has caught up with large models.

4. Our own math: where 1GB comes from, and how fast 3.6B tokens a day is

Three figures computed here from public parameters (recomputed). First, the size math: 1.73B parameters at 4-bit (0.5 bytes each) ≈ 0.87GB, matching the official “about 1GB” claim — that is why it fits on a phone: the model itself takes about a gigabyte, leaving the memory budget for context and the application. Second, the throughput math: 3.6B tokens per day is about 42,000 tokens per second (3.6B / 86,400 seconds) — a vendor peak figure under vendor conditions, to be re-measured at your quantization tier and batch settings before it goes near an SLA. Third, the parameter math: 2.6B to 1.7B is nominally about 35% fewer parameters, but with the previous generation's low base, the comparison that matters is agentic score at equal size, not how many parameters were removed.\n\nTwo selection red lines belong here. First, licensing: the LFM Open License v1.0 is free below $10M in annual revenue and needs a separate arrangement above it — irrelevant for small teams and individuals, mandatory to check for companies at scale before committing. Second, language: 16 languages for the text model, but only 7 EU official languages — check the list for non-English markets before launch, not after.

5. Selection guide: on-device agents, privacy settings, and single-GPU fine-tuning are the three fits

Three fits. First, on-device agent developers: agents with tool calling on phones, laptops, and edge boxes, where data must not leave the device — LFM3-1.7B is one of the few open options where a 256K context, tool-calling scores, and a 1GB footprint hold at the same time. Start with the text model, add VL and Audio only if the use case needs them. Second, privacy and offline settings: in-car, industrial, and medical edge deployments that must run disconnected, where the d1 models' typed decisions (probabilities, choices, numeric estimates in a single forward pass) suit rule-based judgments better than generated text. Third, teams that want to train their own small specialists: single-GPU fine-tuning from one shared base that branches into text, vision, audio, and decision lines saves maintaining a separate model per modality.\n\nThree misfits, stated plainly: frontier general chat quality — use a large-model API, do not expect a 1.7B model to leapfrog; markets outside the language list — wait for coverage; companies already above the $10M revenue line — settle the license before committing. Review scores (preliminary, from public sources, not our own testing): family completeness credibility high (the six-model split, format coverage, and device matrix are consistent across sources); agentic-lead credibility medium (vendor-reported, no independent reproduction, and the 12.7-fold figure includes a low-base effect); on-device deployability credibility high (the size math recomputes cleanly here, and the format and device lists are specific and checkable).

6. Sources and disclaimer

Sources (verified 2026-10-29):\n1) https://www.techtimes.com/articles/329588/20261028/liquid-ai-releases-lfm3-six-open-weight-models-fuse-text-vision-audio-one-2b-backbone.htm — six models on one 1.7B backbone, WebArena-Verified 38.42 vs 3.02, 3.5T-token agentic mid-training;\n2) https://aiweekly.co/alerts/liquid-ai-opens-six-model-lfm3-family-at-256k-context — 22T pre-training, 256K context, about 1GB at 4-bit, 1.8–2.8x faster decoding, B300 at 3.6B tokens per day, tri-hybrid layer counts;\n3) https://www.unite.ai/liquid-ai-launches-lfm3-family-of-six-open-weight-2b-class-models/ — 1.73B parameters, 128K vocabulary, 16 languages, LFM Open License v1.0, day-one partners;\n4) https://www.i-scoop.eu/liquid-ai-lfm3/ — $10M license line, 7 EU official languages, MIT CSAIL background and $250M raise, Mercedes-Benz partnership;\n5) https://www.orcarouter.ai/blog/lfm3-1-7b-base-vs-gemma-4-12b — base published 2026-10-26, BF16 file 3.46GB;\n6) https://local-ai-zone.github.io/blog/October_2026_AI_Model_Updates.html — six-model split, format coverage (BF16/FP8/NVFP4, GGUF, MLX, ONNX), d1 typed-decision definition.\n\nDisclaimer: this site is not affiliated with Liquid AI. This article is product-technology reporting and analysis, not procurement, deployment, or investment advice. Benchmark scores are vendor-reported, marked preliminary, and not yet independently reproduced; license terms follow the official license text — check it before shipping.

The family split: six models, one backbone

Model	Input → output	Context	Use
LFM3-1.7B	Text → text	256K	Text agents, tool calling
LFM3-VL-2B	Text + images + video → text	128K	Reading images, video, documents, screens
LFM3-Audio-2B	Text / audio → text / speech	32K	Speech understanding and generation
LFM3-Omni-2B	Text + images + video + audio → text / speech	32K	Multimodal agents in one model
d1-2B	Text + images → typed answers	128K	Calibrated decisions: probabilities, choices, estimates
d1-Omni-2B	Text + images + audio → typed answers	128K	Typed decisions including audio

WebArena-Verified scores (vendor-reported, preliminary)

★ LFM3-1.7B38.42

LFM2.5-2.6B3.02

数据来源：Sources: Liquid AI release figures (consistent across multiple reports, 2026-10-29); no independent reproduction yet, for order-of-magnitude reference only

Frequently Asked Questions

How do the six LFM3 models relate?

Six branches from one trunk. LFM3-1.7B-Base (1.73B parameters, 22T tokens of pre-training) is the trunk; text, vision, audio, omni, and two decision models are its fine-tuned branches. Only the text model is 1.7B — the rest are 2B class — so do not misread the family's “2B class” label.

How is the 1GB size calculated?

1.73B parameters at 4-bit (0.5 bytes each) ≈ 0.87GB, matching the official “about 1GB” claim; the BF16 file is 3.46GB. 4-bit quantization is what makes phone deployment possible, at the cost of precision loss — measure it on your own tasks before shipping.

Is LFM3 free for commercial use?

Conditionally. The LFM Open License v1.0 is free below $10M in annual revenue and needs a separate arrangement above it. Small teams and individuals are essentially unaffected, but companies at scale must check their revenue position against the official license text first.

Is the 12.7-fold jump from 3.02 to 38.42 believable?

The arithmetic is right (38.42 / 3.02 ≈ 12.7), but watch the base: the previous generation's 3.02 essentially means no agentic ability, so the multiple flatters while the absolute 38.42 is the number for peer comparison. And all of it is vendor-reported, marked preliminary, with no independent reproduction — not a measured conclusion.