LFM3-1.7B needs 1.05 to 1.10 GB for its 4-bit weights, 2 to 3 GB of memory in total at 32K context, and 5 GB or slightly more at its full 256K context. So the full window fits a machine with 8GB of memory, but not a 4GB graphics card at default settings, and memory is not the real cost: Liquid AI’s own figures put a cold 256K prompt on an iPhone 17 Pro at about 16 minutes.
This guide works out what each context length costs, lists the exact files, gives the Ollama, llama.cpp, LM Studio and MLX setup, and explains the US$10 million licence limit. BFWR has not run LFM3. The steps and figures come from Liquid’s model cards and staff replies, the llama.cpp and Ollama source code, and published file listings, and vendor benchmarks are labelled as vendor claims. GB means one billion bytes.
Quick answer
- File to get:
LFM3-1.7B-QAD-Q4_0.gguf(1.05 GB) orLFM3-1.7B-Q4_K_M.gguf(1.10 GB). One file is the whole download. - Memory: about 1.9 GB at 32K, 3.3 GB at 128K and 5.1 GB at 256K, as peak process memory measured by Liquid on an iPhone 17 Pro.
- Ollama:
ollama run lfm3, on v0.40.13 or newer for Linux and Windows. Below 24 GiB of VRAM it starts at 4K context unless you raise it. - llama.cpp: mainline since PR #31252. Use v0.8.0 (build b11960) or newer.
- Good for: local agents, tool calling, structured output and fine-tuning. Poor for: coding agents and questions that depend on what the model remembers.
- Licence: LFM Open License v1.0. Commercial use is free until your organisation reaches US$10 million in annual revenue.
What 256K context really costs
LFM3 is not a plain transformer. Its 24 layers are 13 gated short convolutions, 5 attention layers with a 256-token window, and 6 global attention layers (layer_types in config.json). Only those six keep the whole context, so the usual key-value cache formula applies to them alone.
With num_key_value_heads 4 and head_dim 128, each global layer stores 2 × 4 × 128 values per token. At f16, llama.cpp’s default cache type, that is 2,048 bytes per layer and 12,288 bytes per token across six layers, or 3.22 GB at 262,144 tokens. If all 24 layers held the full context with the same head shape, the cache would be 12.9 GB. Liquid staff confirmed the 12 KiB per token figure on Hugging Face and put the other layers’ state at a few MiB.
| Context | Cache at f16 | Cache at q8_0 | Cache at q4_0 | Liquid’s iPhone 17 Pro peak (Q4_0 weights, f16 cache) |
|---|---|---|---|---|
| 16K | 0.20 GB | 0.11 GB | 0.06 GB | not published |
| 32K | 0.40 GB | 0.21 GB | 0.11 GB | about 1.9 GB |
| 64K | 0.81 GB | 0.43 GB | 0.23 GB | not published |
| 128K | 1.61 GB | 0.86 GB | 0.45 GB | about 3.3 GB |
| 256K | 3.22 GB | 1.71 GB | 0.91 GB | about 5.1 GB |
The cache columns are computed from config.json and ggml’s block sizes. The last column is Liquid’s measurement and grows by about 14 KB per token. Its 32K figures for an M5 Max and a Ryzen AI Max+ 395 are 2.25 and 2.75 GB, so desktops run a little higher.
- 4GB graphics card: the weights plus an f16 256K cache come to 4.27 GB before any working buffers, so the full window does not fit. A 64K window does, on paper. One Hugging Face user reports running the QAD Q4_0 file with Hermes on a 4GB integrated GPU.
- 8GB card or 8GB Mac: 256K fits at about 5 GB. A Mac shares that memory with the system, so 128K, or a q8_0 cache, leaves more room.
- Phones: the card’s line about fitting any phone refers to the 1 GB of weights, not the 256K window.
Time is the larger cost. In the same Liquid test, prefill fell from about 1,050 tokens per second at 32K to 480 at 128K and 270 at 256K, so time to first token went from about 30 seconds to 4.5 minutes to 16 minutes. Liquid has not published the same curve for a desktop. Its advice is to keep the prompt prefix stable and reuse the cache, so each turn processes only new tokens.
For most people 64K to 128K is the sensible setting. Hermes Agent rejects anything under 64,000 tokens. The arithmetic for other models is in our guide to context windows and VRAM.
Files and download sizes
| Build | File or tag | Size | Notes |
|---|---|---|---|
| QAD Q4_0 GGUF | LFM3-1.7B-QAD-Q4_0.gguf |
1.05 GB | Smallest. Trained to recover the accuracy Q4_0 loses, per the card |
| Q4_K_M GGUF | LFM3-1.7B-Q4_K_M.gguf |
1.10 GB | Ollama’s default build |
| Q5_K_M GGUF | LFM3-1.7B-Q5_K_M.gguf |
1.26 GB | |
| NVFP4 GGUF | LFM3-1.7B-NVFP4.gguf |
1.28 GB | Needs llama.cpp b12088 or newer; older builds load it but output garbage |
| Q6_K GGUF | LFM3-1.7B-Q6_K.gguf |
1.43 GB | |
| Q8_0 GGUF | LFM3-1.7B-Q8_0.gguf |
1.85 GB | |
| BF16 GGUF | LFM3-1.7B-BF16.gguf |
3.47 GB | Full precision; the benchmarks below used BF16 |
| MLX 4-bit | LiquidAI/LFM3-1.7B-MLX-4bit |
0.97 GB | Also 5-bit 1.19, 6-bit 1.41, 8-bit 1.84, BF16 3.46 GB |
The model has 1,728,801,536 parameters (Hugging Face API). All text GGUFs sit in LiquidAI/LFM3-1.7B-GGUF.
Ollama setup
Use Ollama v0.40.14 or newer. Linux and Windows support arrived in v0.40.13, and v0.40.14 stopped the OpenAI-compatible endpoints from replacing packaged sampling defaults with a temperature of 1.0. LFM3 ships with 0.1, and agents usually connect through /v1.
ollama run lfm3
ollama run lfm3:1.7b-qad-q4_0
ollama launch hermes --model lfm3
ollama launch openclaw --model lfm3
The library page shows the tools and thinking badges. The tags are 1.7b-qad-q4_0 (1.05 GB), 1.7b-q4_K_M (1.10 GB), 1.7b-q8_0 (1.85 GB) and 1.7b-bf16 (3.47 GB), plus the MLX builds 1.7b-nvfp4, 1.7b-mxfp8 and 1.7b-mlx-bf16.
The 256K badge is the model’s limit, not what you get. Ollama’s default follows VRAM: 4K below 24 GiB, 32K from 24 to 48 GiB, and 256K at 48 GiB or more. Set it yourself on a small machine:
OLLAMA_CONTEXT_LENGTH=64000 ollama serve
ollama ps
The CONTEXT column in ollama ps shows what was allocated. See our Hermes and OpenClaw guides for the agent side.
llama.cpp setup
The card speaks of LFM3-compatible builds, but support is now in mainline llama.cpp. PR #31252 merged on 27 October 2026 and shipped in v0.8.0 (b11960, 28 October). Ollama v0.40.13 bundled b11963. Liquid’s commands:
llama-cli -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M \
--temp 0.1 --top-k 50 --repeat-penalty 1.1
llama-cli -hf LiquidAI/LFM3-1.7B-GGUF \
--hf-file LFM3-1.7B-QAD-Q4_0.gguf \
--temp 0.1 --top-k 50 --repeat-penalty 1.1
For an agent, this follows the server command in Liquid’s agent guide (written for LFM2.5-2.6B), with the model and context changed:
llama-server -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M \
--jinja --host 127.0.0.1 --port 8080 \
-c 65536 -ngl 99 \
--temp 0.1 --top-k 50 --repeat-penalty 1.1
Always set -c: with no value llama.cpp loads the model’s context, and --fit (on by default) may cut it to fit memory, down to 4,096. The default port is 9931. The automatic setting creates four slots sharing one cache, so 256K is a shared pool.
For the full window with less memory, use -c 262144 -ctk q8_0 -ctv q8_0 (OLLAMA_KV_CACHE_TYPE=q8_0 on Ollama). A quantized cache needs Flash Attention, which llama.cpp then enables itself. Liquid has published no quality comparison for this on LFM3.
For maths and science, the card switches to temperature 1.0, top_p 0.95 and top_k 20.
LM Studio and MLX
LM Studio needs llama.cpp engine 2.65.1 or newer, per the GGUF card. Search for LiquidAI, download the Q4_K_M file, then load it in the Developer tab, enable tool use, set the context length and start the server on http://localhost:1234.
On Apple Silicon, the MLX route is MLX-VLM 0.7.11 or later, because no mlx-lm release loads LFM3 yet (PR #2016 was still open on 9 November):
pip install -U "mlx-vlm>=0.7.11"
python -m mlx_vlm.generate --model LiquidAI/LFM3-1.7B-MLX-4bit \
--prompt "What is the capital of Australia?" --enable-thinking \
--temperature 0.1 --top-k 50 --repetition-penalty 1.1 --max-tokens 16384
What Liquid’s benchmarks show
Everything here is Liquid’s own reporting. Speed and memory come from a chart image on the card: Q4_0 GGUFs in llama.cpp at 32,768 tokens of context. LFM3 ran on its own build and the others on b11744.
| Device | LFM3-1.7B | Qwen3.5-4B | Qwen3.5-9B | Gemma 4 E4B |
|---|---|---|---|---|
| Ryzen AI Max+ 395 (Vulkan) | 3,771 / 129.6 / 2.75 GB | 1,355 / 51.2 / 6.00 GB | 1,001 / 34.0 / 8.41 GB | 337 / 37.6 / 9.52 GB |
| Apple M5 Max (Metal) | 8,028 / 237.1 / 2.25 GB | 3,005 / 100.3 / 5.16 GB | 1,927 / 71.3 / 7.80 GB | 3,191 / 86.9 / 6.69 GB |
| iPhone 17 Pro | 1,061 / 43.6 / 1.90 GB | 339 / 15.5 / 4.53 GB | 229 / 10.3 / 6.82 GB | 395 / 15.4 / 6.15 GB |
Each cell is prefill and generation tokens per second, then peak memory. Quality scores were produced with vLLM in BF16, so they describe the 3.47 GB weights rather than the 1 GB file.
| Benchmark (vendor scores) | LFM3-1.7B | Gemma 4 E4B | Qwen3.5-4B | Qwen3.5-9B |
|---|---|---|---|---|
| AutomationBench (public) | 41.29 | 1.92 | 8.67 | 15.10 |
| BFCL V4 (tool use) | 71.39 | 48.25 | 55.12 | 62.83 |
| MCP-Atlas | 32.57 | 6.19 | 41.37 | 42.67 |
| SWE-bench Verified (OpenCode) | 43.40 | 0.50 | 41.20 | 51.20 |
| GPQA Diamond | 63.37 | 57.88 | 75.76 | 80.40 |
| AA-Omniscience accuracy | 7.07 | 9.83 | 15.67 | 13.17 |
| RULER 128K (long context) | 65.56 | 18.82 | 84.63 | 87.11 |
Across the card’s 29 scores, LFM3 beats Gemma 4 E4B on 28, Qwen3.5-4B on 22 and Qwen3.5-9B on 14. The headline average (49.37 against 47.24 for the 9B) depends on the three AA-Omniscience rows; without them the two average 51.20 and 51.24. Read RULER against the 256K claim: the window exists, but both Qwen models score about 20 points higher at 128K.
Where a 1.7B model is the right choice
It suits background work on hardware with little to spare: routing, extraction into a fixed format, calling a known set of tools, and scheduled agents. If you have a 12GB or 16GB card and want a general assistant, compare our Gemma 4 12B and Qwen3.5 9B guide, the small models in our Nemotron 3.5 guide, the 8GB VRAM list and the model directory.
Liquid itself calls the model “not recommended for agentic coding or complex multi-step reasoning”. Its card also says to pair it with retrieval or search for knowledge-heavy work.
LFM3-Omni-2B: audio and vision on the same backbone
LFM3-Omni-2B has 2,166,928,384 parameters and takes text, images, video and 16 kHz audio. It answers in text, 24 kHz speech or a mix. It is a non-reasoning model, and its card gives a context length of 32K, not 256K, although its config says 128,000 and Ollama’s page 125K.
The GGUF is LFM3-Omni-2B-Q4_K_M.gguf (1.10 GB) plus a projector, from 0.29 GB (mmproj-LFM3-Omni-2B-Q4_K_M.gguf) to 0.51 GB (mmproj-LFM3-Omni-2B-Q8_0.gguf). With -hf, llama.cpp fetches the projector for you.
llama serve -hf LiquidAI/LFM3-Omni-2B-GGUF:Q4_K_M --reasoning off --temp 0 --port 8080
llama cli -hf LiquidAI/LFM3-Omni-2B-GGUF:Q4_K_M \
--audio speech.wav -sys "Perform ASR in English." -p "" --temp 0
llama-tts -m LFM3-Omni-2B-Q4_K_M.gguf -mm mmproj-LFM3-Omni-2B-Q8_0.gguf \
-p "Hello world" -o out.wav
llama.cpp handles images, audio and video (video needs ffmpeg), Ollama images and audio, and LM Studio images only, per the card. On Ollama, /api/chat ignores the audio field (issue #19378).
Common failures and fixes
Error: this model requires MLX support, but the MLX runtime is not available: you are on Ollama v0.40.12 on Linux or Windows (#19369). Update to v0.40.13 or newer.error loading model: unknown model architecture: 'lfm3': the llama.cpp build is too old (Ollama #19347). Use v0.8.0 or newer.constexpr function never produces a constant expressionon Metal: a shader failure on macOS 14 and 15 (#31272, #31337). Fixed by #31360, so use v0.9.0. SettingGGML_METAL_BF16_DISABLE=1was the reported workaround.Unexpected empty grammar stack after accepting piece: 6 (30): HTTP 500 whenresponse_formatmeets disabled reasoning (#31328). Fixed by #31374, in v0.9.0.- A stray
</think>in answers: a Liquid team member says leaked reasoning is expected because the model was trained to reason, and settingenable_thinking=Falseis not fully reliable. Strip the tag in your client. - Hermes describes actions instead of taking them: Liquid recommends
hermes config set agent.tool_use_enforcement true.
Security defaults
Ollama binds 127.0.0.1:11434 by default, and llama-server binds 127.0.0.1 with no API key set. Keep them on loopback. Leave file and shell actions on approval in Hermes or OpenClaw.
The licence: free commercial use, up to US$10 million
LFM3-1.7B and LFM3-Omni-2B carry the same LFM Open License v1.0 (Hugging Face shows license:other, so read the file). It follows Apache 2.0, with one added section. Section 1 defines the Threshold as annual revenue of US$10,000,000 or more, and defines Commercial Use broadly as any use for direct or indirect commercial advantage or monetary compensation. Section 5 then has three clauses:
- 5(a): the grant for Commercial Use is conditional on you or your Legal Entity not exceeding the Threshold.
- 5(b): Commercial Use of the Work or a Derivative Work by a Legal Entity above the Threshold is not licensed under the agreement.
- 5(c): the Threshold does not apply when a Qualified Non-Profit Organization (a 501(c)(3) body or foreign equivalent) uses the Work or a Derivative Work for Non-Commercial or Research Purposes.
Individuals and companies under the line can use and ship it commercially at no charge. A company above it needs a separate deal, including for internal tools, since indirect advantage counts.
- Group revenue: Legal Entity includes everything that controls, is controlled by, or is under common control with the acting entity, so a small subsidiary does not get its own count.
- Fine-tunes: Section 5 names Derivative Works, so the limit follows your fine-tune. Section 4 requires you to pass on the licence and mark changed files.
- Termination: Section 11 ends the licence automatically on any breach and requires you to delete all copies.
- Not stated: which year’s revenue counts, and what happens at exactly US$10 million, since the definition includes that figure while Section 5 speaks of exceeding it. The text has no territory or user-count limit and no clause on outputs.
This is a reading of the published text, not legal advice.
FAQ
How much RAM does LFM3 need for 256K context?
About 5 GB in Liquid’s iPhone 17 Pro measurement, with Q4_0 weights and an f16 cache. A q8_0 cache would save about 1.5 GB by calculation.
Does LFM3 work on Ollama on Windows and Linux?
Yes, from v0.40.13. On v0.40.12 it was Apple Silicon only, which caused the MLX runtime error.
Can my company use LFM3 commercially?
Yes, if the whole corporate group is under US$10 million in annual revenue. Above that, commercial use is not licensed.
Related guides and tools
- Hermes Agent with Ollama
- OpenClaw with a local model
- Gemma 4 12B vs Qwen3.5 9B
- NVIDIA Nemotron 3.5 locally
- LLM context windows and VRAM
- Models for 8GB of VRAM
- Local AI model directory
Sources
- LFM3-1.7B model card, config.json and LICENSE
- LFM3-1.7B-GGUF and LFM3-1.7B-MLX-4bit
- LFM3-Omni-2B and LFM3-Omni-2B-GGUF
- Liquid’s long-context memory and prefill figures (Hugging Face discussion 5)
- Liquid on leaked reasoning (discussion 20) and a user report on a 4GB iGPU
- Liquid AI: Introducing LFM3 and Liquid’s agent harness guide
- llama.cpp PR #31252, v0.8.0 release notes, src/models/lfm3.cpp and server README
- llama.cpp issues #31272 and #31328, PRs #31360, #31374 and #31608
- Ollama lfm3 library page, v0.40.13 and v0.40.14 release notes, context length and FAQ
- Ollama issues #19369, #19347, and #19378
- mlx-lm PR #2016 and Hermes Agent provider docs
Updated 9 November 2026. File sizes, parameter counts, licence text, Ollama tags, llama.cpp and Ollama releases and source, and GitHub issues were all read on that date. Speed and quality figures are Liquid AI’s. BFWR has not run this model.
Work out what your hardware can actually run
Free calculators and directories, sized from real model weights — before you download gigabytes or order a card:


Leave a Reply