LFM3 at 2B: Pick the Branch, Skip Agentic Coding, and the 128K Hosted Cliff

Attributed compile + one small offline run on my box

Primary (Liquid AI / Hugging Face, launch 2026-10-27): Build specialized multimodal models with LFM3 (LiquidAI authors on HF), Introducing LFM3 (Liquid AI blog), LFM3 collection.

Secondary: VentureBeat — Carl Franzen, 2026-10-27 for hosted console prices, 128K vs 256K context, license revenue gate, memory footnotes, and the Liquid docs quote against agentic coding.

Benchmarks and on-device speed numbers below are Liquid’s (or VentureBeat’s reading of Liquid’s tables / console) unless I say otherwise. This box did not download multi-GB weights and did not call generate() / system_one. What I ran is a collection-id check, a dry first-call skeleton map, and list-price arithmetic on VentureBeat’s published console rates.

---

Liquid AI shipped LFM3 on 2026-10-27: six ~2B open-weight models branched from one LFM3-1.7B-Base backbone — text, vision-language (with video), audio, omni, and two decision models (d1-2B, d1-Omni-2B). The useful builder move is not “run the biggest generalist.” It is pick the branch first, then decide hosted vs downloadable — because the traps (context, license, coding fit) sit between those choices.

This is distinct from the earlier Liquid d1 open CPU-triage write-up in this series: that post was about decision models as a product shape; this one is the LFM3 family release and how the branches wire into transformers.

Pick the branch before you pick the harness

From the HF / Liquid tables (vendor framing):

Model	Reach for it when…	First transformers entry (HF blog)
LiquidAI/LFM3-1.7B	Local agents, tool calling, structured text; 256K on the downloadable text model	pipeline("text-generation", …)
LiquidAI/LFM3-VL-2B	Images + video → text; on-device visual Q&A	AutoModelForImageTextToText
LiquidAI/LFM3-Audio-2B	Voice chat / ASR / TTS / tool calls from speech (18 languages claimed)	AutoModelForMultimodalLM
LiquidAI/LFM3-Omni-2B	One on-device model that sees, listens, and talks back (~2.2B despite the “2B” name)	AutoModelForMultimodalLM
LiquidAI/d1-2B	Many calibrated decisions in one forward pass over text/images/video	AutoModel + system_one(...)
LiquidAI/d1-Omni-2B	Same decision shape when inputs include audio	(see omni model card; same idea)

Architecture note (Liquid, for intuition only): 24 layers, tri-hybrid mix of 13 gated short convolutions + 5 SWA (256-token window) + 6 global attention; 22T pretrain then 3.5T mid-train that extends context to 256K. Liquid says LFM3-1.7B beats their prior LFM2.5-2.6B with ~35% fewer parameters, with the biggest gains on agentic browser/business tasks. Treat that as vendor positioning until you measure your own loop.

When 1.7B is the wrong tool (coding / knowledge / long-context accuracy)

Liquid’s own table (reproduced and annotated by VentureBeat) is the honest “don’t oversell” slide. Selected Liquid-reported scores:

Eval	LFM3-1.7B	Qwen3.5-9B
WebArena-Verified	38.42	28.94
AutomationBench (public)	41.29	15.10
BFCL V4	71.39	62.83
SWE-bench Verified (OpenCode)	43.40	51.20
GPQA Diamond	63.37	80.40
RULER 128K	65.56	87.11
MCP-Atlas	32.57	42.67

Liquid’s blog says larger models still win on agentic coding and knowledge. VentureBeat quotes Liquid’s documentation more bluntly: not recommended for agentic coding or complex multi-step reasoning due to its small size. That is the rule of thumb for this post: use LFM3-1.7B for narrow local agents, triage, tool routing, and browser/business loops where Liquid’s table shows strength — and do not treat it as a drop-in SWE / GPQA / long-RULER replacement for a ~9B (or larger) coder.

Hosted vs local: the 128K cliff, prices, RAM, license

Three traps that show up before quality:

1. Context. Downloadable LFM3-1.7B is documented at 256K. VentureBeat’s review of the signed-in Liquid console lists 128,000 context for the hosted LFM3 generation entries. If your agent prompt assumes the downloadable window on the API, you hit a 128K hosted cliff.
2. List prices (VentureBeat console, USD / 1M tokens). LFM3-1.7B $0.08 in / $0.02 cached / $0.20 out; VL $0.10 / $0.02 / $0.50; Audio and Omni $0.10 / $0.02 / $1.00 plus audio in $2 / out $10 per 1M audio tokens. Hosted d1.1 is listed at $0.02 in with no output charge — and VB notes that hosted identifier may differ from downloadable d1-2B.
3. “~1GB” is weights, not the app. Liquid’s edge pitch is ~1GB at 4-bit for the text model. That is model storage, not full process RAM. Liquid separately cites ~1.6GB KV cache at 128K; VB’s table of Liquid-measured llama.cpp Q4_0 footprints on Ryzen AI Max+ 395 / Apple M5 Max / iPhone 17 Pro is higher still (e.g. 2.75GB / 2.25GB / 1.90GB reported memory under that specific 32K-in workload). Local decode speed claims (1.8–2.8× the next model in Liquid’s comparison) are Liquid-measured on that hardware — label them as such.
4. License. LFM Open License v1.0: per Liquid’s terms as summarized by VentureBeat, businesses with >$10M annual revenue need a separate commercial agreement. Local weights remove per-token API charges; they do not erase licensing, hardware, or ops cost.

First transformers calls (shapes only — not executed here)

HF’s launch post is the practical card. Text:

from transformers import pipeline  # requires recent transformers; not installed on this box

generator = pipeline(
    "text-generation",
    model="LiquidAI/LFM3-1.7B",
    trust_remote_code=True,
)
messages = [{"role": "user", "content": "What is C. elegans?"}]
# NOT EXECUTED here — would need weights + generate()

VL uses AutoModelForImageTextToText + AutoProcessor (transformers>=5.21.0 per HF). Audio / Omni use AutoModelForMultimodalLM (transformers>=5.20.2). Decision models use AutoModel and system_one(state, questions) — several named questions, one pass, calibrated answers:

from transformers import AutoModel  # skeleton only

# model = AutoModel.from_pretrained("LiquidAI/d1-2B", trust_remote_code=True, ...)
questions = {
    "refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"},
    "team": {
        "type": "choice",
        "instructions": "Which team should handle this?",
        "criteria": {
            "billing": "Charges, refunds, invoices",
            "shipping": "Delivery and parcels",
            "technical": "App or site faults",
        },
    },
    "urgency": {
        "type": "score",
        "instructions": "How urgent is this ticket?",
        "criteria": ["low", "medium", "high"],
    },
}
# out = model.system_one("I was charged twice...", questions)  # NOT EXECUTED

HF also describes newer typed answers (Point / Read / Estimate / Pick) on the decision cards. I did not run those. The one-pass shape is the product point: classify / extract / route without a free-form chat turn when you do not need one.

What I actually ran (offline)

runs/28-lfm3/preflight.py — no weight download. Trimmed output:

transformers: NOT INSTALLED
torch: NOT INSTALLED
Weights: NOT DOWNLOADED (by design; multi-GB)

HF collection LiquidAI/lfm3 — API OK:
  LiquidAI/LFM3-1.7B-Base, LFM3-1.7B, LFM3-VL-2B, LFM3-Audio-2B,
  LFM3-Omni-2B, d1-Omni-2B, d1-2B  (all expected core ids present)

Hosted list-price arithmetic (VB console; NOT a bill):
  LFM3-1.7B   20k in + 2k out, no cache $0.002000 | 80% cache $0.001040
  LFM3-VL-2B  20k in + 2k out, no cache $0.003000 | 80% cache $0.001720
  1M+1M text tokens: 1.7B $0.28 | VL $0.60 | Audio/Omni $1.10
  d1.1 1M in (no out charge): $0.02

Toy agent loop MY ASSUMPTIONS — 40 turns × (20k in + 2k out):
  LFM3-1.7B  no cache $0.0800 | 80% hit $0.0416
  LFM3-VL-2B no cache $0.1200 | 80% hit $0.0688

Dry skeletons for pipeline / AutoModel* / system_one: STATUS NOT EXECUTED

Costs are arithmetic on VentureBeat’s list rates under assumptions I state. They are not a bill, not latency, and not quality.

Checklist before you put LFM3 in a path

1. Pick the branch (text / VL / audio / omni / d1) — do not default to Omni “just in case.”
2. Coding / hard reasoning? Prefer a larger model; Liquid’s docs say so for this size class.
3. Hosted? Budget 128K context and the console rate card; do not assume downloadable 256K.
4. Local? Plan RAM beyond the ~1GB 4-bit weight number; check LFM Open License v1.0 if revenue >$10M.
5. Many structured questions? Prefer d1 system_one over a chat model inventing JSON.
6. Wire an eval on your tickets, invoices, or tool traces. Vendor WebArena wins do not ship your workflow.

Specialize the small model when the job is narrow. Escalate when it is not.

---

YongBo Yu · Toronto AI engineer · https://yongbo-yu.vercel.app · GitHub YongBoYu1