Liquid AI's table puts LFM3-1.7B 2.13 points ahead via one evaluation; licence bars commercial use above $10M revenue

Remove one evaluation and a 2.13 point lead becomes minus 0.04

Liquid AI's launch post for LFM3, dated 27 October 2026, prints a 29 row text benchmark table. Its Average row gives LFM3-1.7B 49.37 and Qwen3.5-9B 47.24, a lead of 2.13 points (derived from the printed figures). All six printed averages reproduce when the 29 rows are averaged without weights, so the arithmetic is sound. The lead, though, has a shape. Three of the 29 rows come from one evaluation, AA-Omniscience, and they favour LFM3-1.7B by 28.50, 40.34 and, on accuracy, minus 6.10 points. Together they add 62.74 points, which is 2.16 points of a 29 row average (derived). Set those three rows aside and the other 26 rows average 51.20 for LFM3-1.7B and 51.24 for Qwen3.5-9B, a lead of minus 0.04 (derived).

That does not show that the model is weak. Several rows in the same table show large leads in instruction following and tool use, and the 1.7 billion parameter model is competing with one that has 9.7 billion (5.7 times as many, derived). It shows that the headline average is carried by a measure that rewards a model for declining to answer, and that the table cannot tell a reader how that trade suits their own work. Liquid AI's own post says "Larger models still win on agentic coding and knowledge", so the post does not claim a clean win; the average is what makes the table look like one.

This briefing reads the launch post (https://www.liquid.ai/blog/lfm3), the seven Hugging Face model cards and repositories, the licence file, the pricing page, the model's config.json, Artificial Analysis's definition of the AA-Omniscience measure, and the vendor's wind turbine cookbook. Everything was read between 08:12 and 08:18 GMT on Wednesday 28 October 2026, the day after launch. Nothing was run: every score in this briefing is Liquid AI's own. The second finding is not about accuracy at all. The post never mentions a licence. The model cards name the LFM Open License v1.0, whose text allows commercial use only below 10 million United States dollars of annual revenue for the whole group of companies a user belongs to. The rest of the briefing covers what the family is, what the memory and speed claims reproduce to, and what a deployer should do before relying on any of it.

The licence: free below 10 million dollars of annual revenue, not licensed above it

"Open-weight" is accurate: anyone can download the weights, which are not gated. Whether "open" carries to commercial use depends on a clause the launch post does not mention. The post, read in full (the saved copy and a second fetch at 08:15 GMT were identical in text), contains no licence line, no revenue threshold and no link to the licence text. Each of the seven repositories carries a licence file, and the cards name it as the LFM Open License v1.0. The Hugging Face tag is license:other with the name lfm1.0. The text is identical in all seven repositories apart from a trailing space and a final newline in one (checked by file hash).

The clauses below are quoted exactly. This is a reading of the licence's text for a technical reader, not legal advice: read it with your counsel before any commercial use. Nothing here says the licence is bad or unusual. It says who it binds, and that the post did not.

Three points follow from the text. First, the group rule: the Legal Entity is a union with its parents, subsidiaries and sister companies, so a small team inside a large group is not measured on its own revenue (a reading of the definition, to confirm with counsel). Second, the exact line: the Threshold is defined as ten million dollars "or more", clause 5 speaks of "not exceeding" and "exceeds", and the pricing page splits its tiers into "< $10M" and "> $10M", so a company at exactly ten million is not clearly placed by any of them (derived observation). Third, there is no acceptable use policy, field of use limit or geographic limit in the licence text (derived by reading all of it). The restriction is the revenue line and the redistribution conditions.

Liquid AI's pricing page, read at about 08:14 GMT, puts the same rule in plain language: the open weight models are "free to use, including commercially" and "You only pay us once your company's annual revenue passes $10 million". Below the line the page says use in production is at no cost. Above it, "contact us for a paid license", with "Pricing is structured around your deployment scale", so no price is published. The page says research, education and non-profit use is free with no revenue limit; the licence text is narrower and more specific (a Qualified Non-Profit, non-commercial or research purposes), and the text is what binds. The post says nothing about what terms its launch partners (AMD, Arm, Intel, NVIDIA, Qualcomm, Vercel, OpenRouter, Ambarella, Mercedes-Benz and Shopify) use, and this briefing does not guess.

One plain sentence on what open weights change: weights that have been downloaded cannot be withdrawn, and Liquid AI's rules for their use bind only through the licence, which for a company above the line means a negotiated agreement. Whether the release meets any formal definition of open source AI was not assessed, because that definition was not read for this briefing.

What the 2.13 points are made of

The chart takes every row of the vendor's table and subtracts the Qwen3.5-9B score from the LFM3-1.7B score. Each benchmark has its own scale (some are percentages, one is an index that runs from minus 100 to plus 100), so the bars are points on different rulers, and an average over them is a convention, not a measurement. The three violet bars are the AA-Omniscience rows. LFM3-1.7B is ahead on 14 rows and behind on 15 against Qwen3.5-9B, and ahead on 22 of 29 against the 4.7 billion parameter Qwen3.5-4B (all derived from the table).

Zoom

Keeping the index row and dropping the accuracy and non-hallucination rows is a fair alternative, because the index is the evaluation's own headline number and the other two are components of it. On that reading the lead is 1.02 points (derived). The point is not that one count is right. It is that three of 29 rows are one evaluation measured three ways, so a 29 row average counts it three times, and the sign of the lead depends on that choice.

What those three rows measure: how often the model declines

Artificial Analysis, which publishes the benchmark, defines its index as rewarding correct answers, penalising hallucinations and having "no penalty for refusing to answer", on a scale from minus 100 to 100, where 0 means as many correct answers as incorrect ones. Its article of 16 November 2025 gives the scoring as +1 for a correct answer, minus 1 for an incorrect one and 0 for an abstention. Its page defines the hallucination rate as incorrect answers divided by all answers that were not correct (incorrect, partial and not attempted). The table's "non-hallucination" row appears to be 100 minus that rate (inference from the numbers, below).

From those definitions the three published LFM3-1.7B numbers fix the rest. Accuracy is 7.07 per cent. The index is correct minus incorrect, so incorrect is 7.07 minus 0.50, or 6.57 per cent of questions (derived). That leaves 100 minus 7.07 minus 6.57, which is 86.36 per cent of questions that were neither right nor wrong: partial answers or questions not attempted (derived). As a check, the hallucination rate that follows is 6.57 divided by 92.93, or 7.07 per cent, so non-hallucination is 92.93, exactly the printed figure (derived). The same two routes agree for five of the six models in the table to within 0.03 points of the index, and for LFM2.5-2.6B to within 0.34.

Read plainly: LFM3-1.7B answered correctly least often of the six models, and chose not to commit on about six questions in seven. Qwen3.5-4B answered correctly most often and was wrong on nearly three questions in four. Neither behaviour is a defect in itself. A support assistant that hands unknown questions to a person is helped by abstaining, and a tool that must produce an answer is not. The model card itself says "Use retrieval or search for knowledge-heavy tasks", and the post says search nearly doubles the Humanity's Last Exam score. The table's own Knowledge rows agree: GPQA Diamond 63.37 against 80.40, and Humanity's Last Exam without tools 7.17 against 16.41.

One more caveat. The table labels the rows "(public)", which matches the public subset Artificial Analysis publishes (a dataset named AA-Omniscience-Public-v1.1 on its page). Artificial Analysis says that across 86 models at the October 2026 launch of that version the public and private subsets correlate at 0.991, with a mean difference of minus 0.2 points. That supports reading the public figure as representative. It does not mean Artificial Analysis ran these models: the post does not say who ran the rows or whether Artificial Analysis's harness was used.

The other 26 rows: real leads, real losses

Fair to Liquid AI: the instruction following and tool use rows are not an artefact of one evaluation. LFM3-1.7B leads Qwen3.5-9B by 24.19 points on IFBench and 8.56 on BFCL V4, and by 26.19 on AutomationBench and 9.48 on WebArena-Verified, the two long browser and business app tasks the post names. The post's sentence, "On agentic tasks, LFM3-1.7B competes with models several times its size", holds for those rows. It does not hold for the rows below, and the post says as much.

The model card is blunt about the other side: "It is not recommended for agentic coding or complex multi-step reasoning due to its small size." The largest loss in the whole table is long context recall: RULER 128K is 65.56 for LFM3-1.7B against 87.11 for Qwen3.5-9B and 84.63 for Qwen3.5-4B. The post advertises a 256K context window, and the table's own long context rows are where it trails most. A separate figure points the same way: the model keeps full context in only 6 of its 24 layers (see the memory section). Whether the cheap cache and the weaker recall are linked is an inference this briefing cannot test, and neither the post nor the card says.

Cost per task cannot be put beside these scores. The post gives no price per million tokens for LFM3, and the pricing page, read at about 08:14 GMT, listed three API models (d1, d1-ultra and LFM2.5-24B-A2B) and no LFM3 model. The table gives no tokens per task, and the card says reasoning is on by default, so a price would not give a cost per task even if one were published. The only efficiency figure from outside Liquid AI is the MacPaw quote in the post, covered next. Running the open weights yourself carries no licence fee below the revenue threshold, but the hardware and the tokens are yours.

Who ran the numbers, and what nobody else has measured

Every benchmark in the post is Liquid AI's. The text model card describes how each model was served, which reads as a description of Liquid AI's own runs (inference; it names no outside party), and the omni card says so outright: "All scores were measured by Liquid AI." The one number from outside the company is a quote from a staff engineer at MacPaw. The quotes from AMD, Ambarella, Qualcomm and Mercedes-Benz are statements by launch partners about their plans and intentions, not tests.

Where an independent evaluator would show up, nothing did. Artificial Analysis's page for Liquid AI as a creator lists 10 models, with no scores shown, and the visible names are LFM2.5 models; its sitemap holds no LFM3 page, and its provider page for Liquid AI answered 404. Vals AI's index page, read in full as served, mentions no LFM model. These checks were made between 08:14 and 08:16 GMT on 28 October 2026. A model announced the day before is rarely listed, so the absence is not evidence about quality; it means every score in this briefing is a vendor score.

The vision claims are checkable in more detail than the post allows, because the vision card prints all 62 scores. Averaging them reproduces the headline: LFM3-VL-2B 71.39, Gemma 4 12B 71.74, GPT-5.6 Luna 73.43, Gemini 3.5 Flash Lite 70.02 and Qwen 3.5 2B 64.01 (derived), which round to the printed 71.4, 71.7, 73.4, 70.0 and 64.0. The post's "at least 7.4 points" ahead of similar-size models is exactly 71.4 minus 64.0. The vision card says "at least 8 points", which its own table does not support for the 2.3B Qwen model (7.4, derived).

Six labels that do the work of evidence

Seven checkpoints, one parent, five characters

The post says LFM3 brings "text, vision, audio, omni, and decision models onto one new shared backbone", with six 2B-class models "trained from LFM3-1.7B-Base" and "all branched from the same LFM3-1.7B-Base backbone". The Hugging Face metadata supports the chain for four of the six: the text and vision models name the Base as their parent, d1-2B names LFM3-VL-2B, and d1-Omni-2B names LFM3-Omni-2B. The audio and omni repositories set no parent tag. That is a metadata gap, not evidence against the post: their cards simply describe an LFM3 backbone and do not name the Base. The vision card says it "starts from LFM3-1.7B-Base".

Zoom

Zoom

Two things to read from the table. The 256K context applies to the text model, as the post's wording says ("256K context window for the text model"); audio and omni are 32K and the rest 128K. And "2B-class" covers 1.73 to 2.17 billion parameters (derived from the counts). The post calls the text model 1.7B and the card calls it 1.73B, the same figure. On the day, the text model had 6,311 downloads and 242 to 244 likes on Hugging Face (the interface's counters, definition not checked); the other repositories were far lower. The repository creation dates (the audio repository 8 September, the Base 26 October) are not release dates, and this briefing uses the post's date, 27 October.

Vision, audio and decision claims against the cards

On vision, the vendor's sentence is "matches the performance of Gemma 4 12B" (71.4 against 71.7), "surpasses Gemini 3.5 Flash Lite" (71.4 against 70.0) and "within 2 points of GPT-5.6 Luna" (73.4 against 71.4, exactly 2.0). The table is also clear about where the small model trails the closed ones: Video 61.2 against 72.8 for Luna and 70.0 for Flash Lite, and Agentic 42.5 against 56.1 and 62.4 (differences of 11.6 and 8.8, and 13.6 and 19.9 points, derived). The card's table adds one column the post's omits, Qwen3.5-4B, which averages 73.4 and is 2.0 points ahead at 4.7 billion parameters, about 2.5 times the size (derived). The post's selection is for similar-size models, which is a reasonable choice, but a reader should see the omitted column.

A name check. The vision table and its sentence say "GPT-5.6 Luna". The wind turbine section of the same post says "GPT-6 Luna" twice, and the vendor's cookbook README says "GPT-6 Luna" as well. The post does not say whether these are the same model or one is a typo, and this briefing does not resolve it. The card says medium reasoning effort for the model it calls GPT-5.6 Luna; the cookbook does not name a setting.

The decision models are the least checkable claim, because the headline index is private: the post's "Our Meta index (117 private and public benchmarks)" moves from 58.8 for d1-3B to 63.2 for d1-2B, and the cards call it a board of "over a hundred public and private benchmarks". The public rows move too (Decision Bench 83.6 to 85.2, Decision Index 49.0 to 51.1, ImajevBench 64.0 to 75.5), and the cards list per-benchmark links, but the vendor ran them. The d1-3B parameter count was not read, so the "third fewer parameters" claim is not checked here. The card limits are specific: answers "get less reliable past about 64K tokens of state", knowledge-heavy questions "favor larger models", and the model "is not a chat model".

On spoken against written, the post says d1-Omni-2B, "Asked by voice instead of in writing, it keeps nearly all of its accuracy", and shows three rows: DecisionBench 83.4 written and 78.2 spoken (93.8 per cent of the written score, derived), Typed Decisions 50.5 against 48.0, and JevBench 87.6 against 89.0. The model card's audio table has ten rows, scored as "effective accuracy, where 0 is chance and 100 is every answer right". The other rows show larger drops.

So "nearly all" is true of the three rows chosen and of the held-out spoken questions, and false of spoken short states (heard is 54 per cent of written, derived), Speech-MASSIVE (74 per cent) and Belebele (50 per cent). The card does say this itself: "Long spoken passages lose much of their accuracy, most of all outside Latin-script languages." The post's sentence is the summary, and the card is where the qualification lives. The card's DecisionBench written figure is 83.4, while its benchmark table lists 76.4 for the same named benchmark; the card does not explain the difference in that row, and a subset is the likely reason (inference). The d1-Omni-2B card also says d1-2B "is better calibrated", which implies a calibration difference but gives no figure.

Memory and speed: what reproduces and what does not

The model's config.json lists its layer types in order, which the post does not. Of 24 layers, 13 are gated short convolutions, 5 are sliding window attention with a 256 token window and 6 are global attention, and the six global layers sit at positions 3, 7, 11, 15, 19 and 23. Only global layers keep a cache that grows with the whole context. With 4 key value heads, a head size of 128 and BF16 (2 bytes), each global layer holds 2 times 4 times 128 times 2, or 2,048 bytes per token, so six layers hold 12,288 bytes per token (derived). At 131,072 tokens that is 1.61 GB, and at 128,000 tokens 1.57 GB (derived); the sliding layers add 2.6 MB. The post's "1.6 GB of KV cache" for a full 128K context therefore reproduces from the config (derived, assuming a runtime that keeps only the window for sliding layers and a BF16 cache). At the full 262,144 tokens the same sum gives 3.22 GB (derived).

Zoom

The rivals' figures, 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B at 128K, were not reproduced, because those models' configurations were not read. The ratios are 2.7 and 3.5 (derived from the post's rounded numbers). Weights follow from the parameter count of 1.7288 billion: about 3.46 GB at BF16, 1.73 GB at FP8 and 0.86 GB at 4 bit before overhead (derived), consistent with the post's "about 1 GB at 4-bit". The card's peak memory at 32K context on three devices, which includes the cache and the runtime, is 2.75 GB (AMD Ryzen AI Max+ 395), 2.25 GB (Apple M5 Max) and 1.90 GB (iPhone 17 Pro), read from the figure below.

Zoom

The decode claim, "1.8 to 2.8 times faster than the next one", reproduces from the figure's printed labels: 129.6 against 70.1 on the AMD device (1.85), 237.1 against 106.0 on the Apple device (2.24) and 43.6 against 15.8 on the iPhone (2.76), all derived. The card states the setup: Q4_0 GGUF files, llama.cpp, 32,768 tokens of context, batch size 1, 100 output tokens, GPU offload. It does not show how a 4 bit file scores on any benchmark in the post, and the benchmark tables above were run in BF16 on vLLM, so the speed claim and the quality claim are measured on different artefacts.

Zoom

The throughput claim, 3.6 billion tokens per day on one B300 at 1,024 concurrent requests, is 41,667 tokens per second (derived: 3.6 billion divided by 86,400 seconds), which matches the end of the chart's line at about 41.7 thousand. The workload is stated on the chart and the card: one B300, BF16, SGLang 0.5.22, 1,024 input and 256 output tokens, mean of 3 runs. For LFM3-VL-2B, "over 12,000 tokens per second" uses the same machine with one 512 by 512 image and 1,024 text tokens per request, up to 256 generated. A cost per token cannot be built from this, because the card gives no price for the B300. The 300,000 agents run on the Enron archive has no result, cost or duration. And the claimed 25 per cent speed advantage over LFM2-1.2B at 8K context is a measurement on Liquid AI's own devices that the post does not tabulate.

The specialisation demos: what the cookbook adds

The post's pitch is specialisation: fine-tune the small model on your own data. Its wind turbine example is the best documented, because Liquid AI published the cookbook. The post says LFM3-VL-2B was fine-tuned on about 3,000 public photos for a few hours on one A100 and tested on about 1,400 held-out photos, against the un-tuned model and GPT-6 Luna. Out of the box both identify the defect types in fewer than 10 per cent of images; after tuning, classification accuracy reaches 67 per cent and detection F1 rises from 0.003 to 0.11, against 0.01 for GPT-6 Luna. The cookbook README, read at raw.githubusercontent.com, gives the detail in a table on 1,372 held-out photos.

Fair reading: a small model tuned on a narrow task beats an untuned frontier model that was asked to do it cold, which is what specialisation claims. The README is open about the limits. The tuned model "finds 190 of the 2,030 annotated defects with the right type and box", a recall of 9.4 per cent (derived), and "the dataset only has defective blades", so false alarms on healthy blades were not measured. It keeps each original photo and its augmented copies in the same split, which guards against near-identical images on both sides. It does not say whether the held-out photos come from the same turbines or sites as the training photos, so how far the result transfers to a new wind farm is not established. The README describes the frontier model as run through an API with the prompt held fixed, and does not describe it as fine-tuned; it also does not report a prompt tuned for that model.

The robotics section is labelled a research extension, with stated results: 96.2 per cent average over four LIBERO suites, 45.7 per cent on RoboCasa-GR1, and navigation success of 53 per cent on R2R and 43 per cent on RxR under a strict protocol, with Jetson AGX Orin timings. The post does not say whether those checkpoints are released. The Hugging Face collection for LFM3 lists exactly seven repositories, and a search of the LiquidAI organisation on Hugging Face for LFM3 returned 33 repositories, the five LFM3 checkpoints plus 28 quantised or converted variants; none is a robotics checkpoint. So the robotics numbers describe something a reader cannot download from those listings. The post states the compute for them (about six GPU-hours for the orchestration adapter, 72 for navigation and 170 for the manipulation model, on AMD Instinct MI325X GPUs), which makes them reproducible in principle only with the data and code, neither of which was found.

Stated and not stated: the record, line by line

"Not stated" means not found in the post or the seven cards, read in full. It does not mean the information does not exist.

Technical details

The artifacts below are copied from the saved pages and files. Each caption names the source and the date it was read. Nothing here is exploit code. They are the licence clauses, the settings and files a deployer touches, and arithmetic this briefing derived, marked as such. Because the Python quick start runs code from the repository at load time, a deployer should read that code first (the checklist below says how).

The licence clauses that set the revenue line. Where they live: the LICENSE file in each repository. What a deployer does with them: compare the group's annual revenue with the Threshold before any commercial use, with counsel.

"Threshold" shall mean annual revenue of 10 million United States dollars ($10,000,000) or more.

"Legal Entity" shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, "control" means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity.

"Commercial Use" shall mean any use of the Work for direct or indirect commercial advantage or monetary compensation.

5. Commercial Use Limitation.

(a) The rights granted under this License for Commercial Use are conditioned upon You or Your Legal Entity not exceeding the Threshold.

(b) Any Commercial Use of the Work or a Derivative Work by a Legal Entity that exceeds the Threshold is not licensed under this Agreement.

(c) The Threshold shall not apply to a Qualified Non-Profit Organization's use of the Work or a Derivative Work for Non-Commercial or Research Purposes.

11. Termination. This License will terminate automatically and immediately if You fail to comply with any of its terms and conditions. Upon termination, You must cease all use of the Work and any Derivative Works and delete all copies in Your possession.

The metadata that labels it. Where it lives: the YAML header of each model card. What a deployer does with it: do not rely on the tag; the licence file governs.

library_name: transformers
license: other
license_name: lfm1.0
license_link: LICENSE
pipeline_tag: text-generation
base_model: LiquidAI/LFM3-1.7B-Base

Recommended sampling and the quick start. Where they live: the model card for the text model. What a deployer does with them: use the card's settings, and note that the card's own benchmarks used temperature 0.6 and top_p 0.95 unless a row says otherwise. The call passes trust_remote_code=True, which executes Python from the repository.

General (chat, tool use, agents, coding): temperature=0.1,top_k=50, repetition_penalty=1.1
Math & STEM (math, science): temperature=1.0, top_p=0.95, top_k=20

from transformers import pipeline

generator = pipeline(
    "text-generation",
    model="LiquidAI/LFM3-1.7B",
    trust_remote_code=True,
)

messages = [{"role": "user", "content": "What is C. elegans?"}]

output = generator(
    messages,
    do_sample=True,
    temperature=0.1,
    top_k=50,
    repetition_penalty=1.1,
    max_new_tokens=2048,
)
print(output[0]["generated_text"][-1]["content"])

The chat template and the tool call shape. Where they live: chat_template.jinja and the card. What a deployer does with them: apply the template through the tokenizer rather than by hand, and parse generated tool calls as XML. Reasoning is on by default.

<|startoftext|><|im_start|>system
You are a helpful assistant trained by Liquid AI.<|im_end|>
<|im_start|>user
Explain how a heat pump works.<|im_end|>
<|im_start|>assistant
<think>

<tool_call>
<function=get_weather>
<parameter=city>
Paris
</parameter>
</function>
</tool_call>

The layer order and the numbers the cache arithmetic uses. Where they live: config.json. What a deployer does with them: size the cache for their context length.

"conv_L_cache": 7,
"dtype": "bfloat16",
"head_dim": 128,
"hidden_size": 2048,
"max_position_embeddings": 262144,
"num_attention_heads": 16,
"num_hidden_layers": 24,
"num_key_value_heads": 4,
"sliding_window": 256,
"use_sandwich_norm": true,
"vocab_size": 128000,
"layer_types": [
    "conv",
    "conv",
    "full_attention",
    "conv",
    "sliding_attention",
    "conv",
    "full_attention",
    "conv",
    "sliding_attention",
    "conv",
    "full_attention",
    "conv",
    "sliding_attention",
    "conv",
    "full_attention",
    "conv",
    "sliding_attention",
    "conv",
    "full_attention",
    "conv",
    "sliding_attention",
    "conv",
    "full_attention",
    "conv"
]

The arithmetic. Derived by this briefing from the files above and the post. It assumes a runtime that keeps only the window for sliding layers and a BF16 cache, and ignores the small fixed state of the convolution layers.

KV bytes per global layer per token = 2 (K and V) x 4 KV heads x 128 head_dim x 2 bytes = 2,048
6 global layers: 6 x 2,048 = 12,288 bytes per token
at 131,072 tokens: 12,288 x 131,072 = 1,610,612,736 bytes = 1.61 GB (decimal)
5 sliding layers: 5 x 2,048 x 256 = 2,621,440 bytes = 2.6 MB
at 262,144 tokens: 3.22 GB. Vendor says 1.6 GB at 128K.

weights = 1,728,801,536 parameters x bytes per parameter
BF16 (2): 3.46 GB   FP8 (1): 1.73 GB   4 bit (0.5): 0.86 GB   before format overhead

3.6e9 tokens per day / 86,400 seconds = 41,667 tokens per second

AA-Omniscience, LFM3-1.7B: accuracy 7.07, index 0.50
incorrect = 7.07 - 0.50 = 6.57 percent of questions
neither correct nor incorrect = 100 - 7.07 - 6.57 = 86.36 percent
hallucination rate = 6.57 / 92.93 = 7.07 percent, non-hallucination 92.93 (matches the table)

The repositories and the revisions read. Where they live: Hugging Face. What a deployer does with them: pin the revision hash (the commit) they tested, and verify the downloaded file hashes against the hub. The hashes below were current on the morning of 28 October and will change when the authors push.

LiquidAI/LFM3-1.7B-Base    545fc8f4c7b5fa9f405105274cd29c4f8f44b526
LiquidAI/LFM3-1.7B         40ebe0aa6662c9d40389c2f5d565b6a2c6aebacd
LiquidAI/LFM3-VL-2B        f83d2b0a2296aaa277f394990a9b8b949ddf9691
LiquidAI/LFM3-Audio-2B     9d49701c372d6011d900f6abb40c6b296fb706dd
LiquidAI/LFM3-Omni-2B      ce2458cbce4ad970f92032ed94e3d18d30b04f9c
LiquidAI/d1-2B             cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41
LiquidAI/d1-Omni-2B        c8d48c08dada0191c6b362820f055cfb49c99426

What a deployer should do, in the order worth doing

What could not be verified

• Everything measured. No benchmark, speed or memory number was run or independently measured. Where a figure reproduces, it reproduces from the vendor's own tables or files.
• Rival models' cache sizes. The 4.3 GB and 5.6 GB figures were not reproduced because those models' configurations were not read.
• GPT-5.6 Luna against GPT-6 Luna. Both names appear in Liquid AI's materials; whether they are the same model was not established.
• The d1-3B parameter count, so "a third fewer parameters" is unchecked, and the private decision index.
• Robotics checkpoints. Not found in the collection or an organisation search; whether they exist elsewhere is unknown.
• Pricing. The pricing page lists no LFM3 model; it may be stale or load its listing on the client side, which was not checked.
• The licence's legal effect. This briefing reads the text and does not interpret it. The Open Source Initiative's definition was not read.
• The GitHub cookbook page. GitHub's robots.txt disallows its tree and raw paths for all crawlers, so the cookbook README was read as a single file from raw.githubusercontent.com, whose robots.txt answered 404 (no rules); its other files and notebooks were not read.
• Time-stamped. Read between 08:12 and 08:18 GMT on 28 October 2026. Download and like counts, repository revisions, the pricing page and Artificial Analysis's listings can change within hours.

The question that exposes the gap

If a model's average lead depends on how often it declines to answer, which of your tasks rewards declining, and which punishes it?

Key facts