TL;DR: Liquid AI released LFM3 on October 27, 2026 — six open-weight models around 2B parameters (text, vision, audio, omni, and two decision models) sharing one backbone. The text model’s Q4_K_M GGUF is a 1.10GB file that decodes a vendor-reported 237 tok/s on an M5 Max and 43.6 tok/s on an iPhone 17 Pro. Any hardware you bought in the last five years already runs it; the catches are the license and what a 1.7B model still can’t do.

	LFM3-1.7B	Qwen3.5-9B	LFM2.5-8B-A1B
Best for	On-device agents, 256K context on a phone or NPU-less laptop	Best raw quality in the small-model class	MoE speed with more total knowledge
Size / cost	1.73B params, Q4_K_M 1.10GB — runs on hardware you own	~5.5GB at Q4 — any 8GB GPU	8.3B total/1.5B active, ~4.7GB Q4
The catch	Vendor says don’t use it for agentic coding; $10M revenue license cap	Liquid’s tests show it 1.8× slower on the same device	Same $10M LFM license cap

Honest take: LFM3 is the rare model release where the correct hardware purchase is nothing. If your machine has 4GB of free RAM, you can run the whole six-model family tonight via llama.cpp — and if you were about to buy a GPU for small-model agents, this release is an argument to buy less GPU, not more. Spend money only when you want bigger models resident alongside it.

Two days ago Liquid AI shipped LFM3, and unlike most of what has crossed this site’s desk in October — MiniMax M3.1’s 143.9GB 2-bit file, Mistral Large 4’s still-unreleased 1T weights — this one runs on the laptop you’re reading this on. That makes the hardware question different from usual. It’s not “which GPU do you need” but “what does a good 2B-class family change about what you should buy at all.” Here’s what shipped, the numbers that survived verification, and the two catches the launch posts skip.

What shipped on October 27

Six open-weight models, all built on the same LFM3-1.7B-Base trunk (Liquid AI announcement, October 27, 2026):

Model	Parameters	What it does
LFM3-1.7B	1.73B	Text: chat, RAG, tool calls. 256K context, 128K vocabulary, 16 languages
LFM3-VL-2B	1.92B (1.73B LM + 164M vision encoder + 25M projector)	Image and video understanding
LFM3-Audio-2B	~2B class	Speech in/out: voice chat, ASR, TTS, translation across 18 languages
LFM3-Omni-2B	2.2B	Text + vision + audio in one model
LFM3-d1-2B	~2B class	Decision model: typed outputs with calibrated probabilities instead of free text
LFM3-d1-Omni-2B	~2B class	Decision model with multimodal input

The architecture is what Liquid calls tri-hybrid: 24 layers mixing 13 gated short-convolution layers (kernel size 7), 5 sliding-window attention layers, and 6 global attention layers. The practical consequence of replacing most attention with convolutions is a small KV cache — which is how the 1.7B model holds a 32K-token context in a reported 1.90GB of total memory on an iPhone 17 Pro, weights included. A pure-transformer model of the same size can’t do that.

If you followed our LFM2.5-8B-A1B coverage or the 24B-A2B agent model, the lineage is straightforward: LFM2.5 scaled the family up toward MoE; LFM3 restarts the generation at 2B with multimodal encoders bolted onto one shared trunk. Liquid’s own comparison says the vision model jumped from a 57.6 average to 71.4 across 62 vision benchmarks versus LFM2.5-VL-3B — vendor-run, but a large claimed gap.

The license: free until you make $10 million

Every LFM3 model ships under the LFM Open License v1.0 — Apache 2.0-derived, but with a line Apache doesn’t have: commercial use is free only while your company’s annual revenue is under $10 million USD. At or above that, you need a paid commercial license, and the threshold follows modified models and fine-tunes too (LFM Open License User Guide).

For a home lab this is a non-issue — personal and research use is unrestricted. But it’s not open source by OSI’s definition, and a Hugging Face discussion on the LFM license makes exactly that argument. If you’re building something at work on top of LFM3, flag the revenue clause to whoever owns legal risk before it ships. This is the same license family our sister site covered in its LFM2.5 family review — nothing changed in v1.0’s substance, only which models carry it.

The speed claims, run through the bandwidth formula

Liquid’s headline: at 32K context, LFM3-1.7B prefills at least 2.5× faster and decodes at least 1.8× faster than every compared model on all three of its test devices (AMD Ryzen AI Max+ 395, Apple M5 Max, iPhone 17 Pro), at the lowest memory use. Test setup: llama.cpp, Q4_0 GGUF, batch size 1, 100 output tokens.

The absolute numbers, and our standard sanity check — decode speed can’t exceed memory bandwidth divided by bytes read per token (~1.1GB for this file):

Device	Bandwidth	Ceiling (tok/s)	Reported decode	Plausible?
Apple M5 Max	614 GB/s	~558	237.1 tok/s	Yes (42% of ceiling)
AMD Ryzen AI Max+ 395	256 GB/s	~233	129.6 tok/s	Yes (56%)
iPhone 17 Pro	—	—	43.6 tok/s, 1,061 tok/s prefill	Implies ~48 GB/s of effective reads — phone-plausible

The M5 Max and Ryzen figures come from a third-party setup write-up reporting Liquid’s table; the iPhone row is from the model card. All pass the ceiling check — note that models this small land below the usual 70–95% of the bandwidth bound, because per-layer launch overhead starts to dominate when each token only reads a gigabyte. Separately, Arm reported provisional numbers with SME2 enabled: over 400 tok/s prefill and over 50 tok/s decode on two CPU threads. Treat every number in this section as vendor-adjacent — as of October 28 no independent harness had reproduced them.

What should your own hardware do? Scale by bandwidth. A used RTX 3090’s 936 GB/s puts the theoretical ceiling near 850 tok/s — in practice expect overhead to cap it far lower, but “faster than you can read by an order of magnitude” is a safe bet on any discrete GPU. A DDR5 dual-channel desktop (~80 GB/s) has a ~73 tok/s ceiling, which is why this model class is the one place CPU-only inference genuinely works. Check your own card and model combination in our VRAM calculator.

The benchmark fine print

Liquid reports LFM3-1.7B averaging 49.37 across 29 benchmarks against 47.24 for Qwen3.5-9B — a model with roughly five times the parameters. Two things to know before repeating that:

The lead is thin and uneven. It’s 2.13 points on an average of Liquid’s chosen 29 benchmarks. One independent briefing points out the margin leans heavily on a single evaluation. Where the gap is real and large is the agent benchmarks: LiveClaw 45.60 vs 35.77 and AutomationBench 41.29 vs 15.10 against Qwen3.5-9B. Tool-calling is visibly what this model was trained for.

Liquid itself tells you what not to use it for. The model card says LFM3-1.7B is not recommended for agentic coding or complex multi-step reasoning. A 1.7B model that beats a 9B on curated averages is still a 1.7B model when you ask it to refactor a codebase. For local coding you’re still in Nemotron Bolt E4B territory at minimum, and realistically a 12GB-card class model from our 12GB VRAM guide.

Where LFM3 fits: always-on, low-power jobs — voice control, image triage, structured routing decisions, document Q&A over the 256K window — on hardware that would be embarrassing to dedicate a real GPU to.

How to run it today (and the error you’ll hit)

llama.cpp works now; Ollama is mid-transition. The weights are on Hugging Face as LiquidAI/LFM3-1.7B-GGUF (plus MLX, ONNX, FP8, and NVFP4 variants), and the card requires llama.cpp b11960 (v0.8.0) or newer:

llama-cli -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M
# first run downloads the 1.10GB Q4_K_M file, then:
# > hi
# Hello! How can I help you today?

On Ollama, here’s the problem we verified — pulling the GGUF on v0.40.12 or older fails:

$ ollama run hf.co/LiquidAI/LFM3-1.7B-GGUF:Q4_K_M
Error: error loading model: unknown model architecture: 'lfm3'

That’s ollama/ollama issue #19347, opened October 27 and closed the next day: GGUF support rides on Ollama’s bundled llama.cpp, and the fix (PR #19344) lands in the release after v0.40.12. Two working paths in the meantime: use llama.cpp directly as above, or — on a Mac only — ollama run lfm3, which pulls the MLX build from Ollama’s registry. That registry tag is Apple-only right now, which confused enough people to generate its own issue. If you’re on Windows or Linux, llama.cpp is tonight’s answer; check ollama -v before blaming the model file.

The quantization menu is shallow but sufficient: community conversions run from 0.92GB (IQ3_S) to 1.85GB (Q8_0). At these sizes there’s no reason to go below Q4_K_M — the 180MB you save costs measurable quality and no meaningful speed.

What hardware you actually need

Honestly: whatever you have. This is the audio.cpp situation again — the models are small enough that the hardware question inverts. A 1.10GB weights file plus a conv-dominated cache fits in:

• Any phone from the last ~3 years. The iPhone 17 Pro number above is the vendor’s own, but 1.90GB total memory at 32K context leaves headroom on anything with 8GB of RAM.
• Any laptop, no GPU required. Arm’s two-thread numbers and the ~73 tok/s DDR5 ceiling both say CPU inference is usable — for a 2B model, reading speed (~10 tok/s) stopped being the bottleneck years ago.
• A Raspberry Pi 5, in the same way LFM2.5 ran on one: it works, at single-digit-to-low-double-digit tok/s, and the Pi’s lack of SME2 means none of Arm’s headline speedups apply. Fine for a voice box, frustrating for anything interactive.
• Any discrete GPU, where the real question is what else the card is for.

That last point is the only purchase decision in this story. A 2B model doesn’t justify buying anything — but if LFM3 is the always-on layer of a home stack (voice in, vision triage, decision routing) and you want a 20–30B-class model resident behind it for the heavy questions, that’s a 12GB-to-24GB card decision, and it’s the same decision it was last week.

What to actually buy

Prices as of October 2026, from the site’s verified price table:

Your situation	The machine	Price	Where
You own any computer or recent phone	Nothing — run LFM3 tonight	$0	—
Always-on small-model box with headroom for 7–13B models behind LFM3	Used RTX 3060 12GB	$260–$296	Check price
You want LFM3 as the front end and 24GB-class models resident behind it	Used RTX 3090 24GB	$1,456–$1,558	Check price
Want to test the stack before spending anything	Rented 3090, from $0.07/hr	pay per hour	Vast.ai

The first row is the honest default. The used-3090 row exists because “small model handles 90% of requests, big model handles the rest” is the architecture these decision models are clearly built for — d1-2B returning a typed, calibrated answer about whether to escalate to the big model is the cheapest router you can run.

FAQ

Is LFM3 actually open source? No. The LFM Open License v1.0 is Apache-based but caps free commercial use at $10M annual company revenue, which fails OSI’s definition. For personal, research, and small-business use it behaves like a permissive license.

Can I run LFM3 in Ollama on Windows or Linux today? Only once your Ollama build is newer than v0.40.12 — older builds fail with unknown model architecture: 'lfm3'. Until the fix reaches your platform, use llama.cpp b11960+ directly. The ollama run lfm3 registry tag is Apple-only (MLX runtime).

Does the 256K context really work on small hardware? The architecture supports it, and the conv-heavy layer mix keeps the cache small — but the vendor’s own speed tests were run at 32K. Expect prefill time, not memory, to be what you feel at six-figure contexts on a CPU or phone.

Is LFM3-1.7B better than Qwen3.5-9B? On Liquid’s 29-benchmark average, by 2.13 points — a thin, vendor-selected margin. On the agent/tool-calling benchmarks the gap is large and in LFM3’s favor; for coding and multi-step reasoning, Liquid itself says to use something bigger.

Should I buy a GPU for LFM3? Not for LFM3 alone. Buy (or keep) a GPU for the larger models you’ll run next to it, and size that choice with the VRAM calculator.

Recommended Gear

• RTX 3060 12GB (used) — $260–$296; the cheapest card worth pairing with a small-model front end
• RTX 3090 24GB (used) — $1,456–$1,558; 936 GB/s and room for a 27B-class model behind LFM3

Sources

• Introducing LFM3: Every Modality for Specialized Intelligence — Liquid AI
• LiquidAI/LFM3-1.7B model card — Hugging Face
• LiquidAI/LFM3-1.7B-GGUF — Hugging Face
• LFM Open License User Guide — Liquid AI Docs
• Ollama issue #19347: unknown model architecture ‘lfm3’ — GitHub
• Ollama issue #19369: lfm3 registry tag requires MLX runtime — GitHub
• Arm on LFM3 SME2 CPU performance (provisional) — Arm/X
• How to Run Liquid AI LFM3 Locally: Setup & Benchmarks — five.reviews
• Liquid AI Launches LFM3 Family of Six Open-Weight 2B-Class Models — Unite.AI
• LFM3’s 2.13-point lead rests on one evaluation — PK Sharma Briefing
• Liquid AI’s new LFM3 models put browser agents, voice commands, and image recognition onto local devices — VentureBeat

Last updated October 29, 2026. Prices and specs change; verify current rates before purchasing.

Get the numbers before you buy

New GPU and mini-PC benchmarks, VRAM thresholds, and price checks — sent when there's something worth acting on, not on a schedule. No spam, unsubscribe anytime.

Email address Subscribe