Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance.
LFM3-VL-2B
LiquidAI·Instruct·released 19 Oct 2026
ollama run lfm3-vl:2bLiquidAI's LFM3-VL-2B is an instruct vision-language model with 1.9B parameters, built on the LiquidAI/LFM3-1.7B-Base model. It accepts text and images as input and produces text as output, with a context length of 128,000 tokens. It is released under the lfm1.0 licence, which allows limited commercial use. The page below covers its VRAM needs and how to run it locally.
How much VRAM does LFM3-VL-2B need?
At Q4_K_M LFM3-VL-2B needs about 2.0 GB with an 8K context (2.0 GB at Q4, 4.6 GB at full precision). Pick your GPU to see what fits and how fast it runs.
Needs: ≈1.9 GB · Q4_K_M · 8K context — Pick a GPU to check the fit.
Advanced: context 8K · KV cache 16-bit
Smallest GPU for each quant and context (KV cache 16-bit)
| Quantisation | 4K | 8K | 16K | 32K | 64K |
|---|---|---|---|---|---|
| Q3_K_M | |||||
| Q4_K_M | |||||
| Q5_K_M | |||||
| Q6_K | |||||
| Q8_0 | |||||
| FP16 |
≤16 GB card24–32 GB48–96 GBmore
| Quantisation | Total |
|---|---|
| Q2_KHeavy lossSmallest, noticeably worse quality | ≈1.6 GB |
| Q3_K_MNoticeable lossSmall, some quality loss | ≈1.7 GB |
| Q4_K_MrecommendedSmall lossRecommended default for local use | ≈1.9 GB |
| Q5_K_MNear-losslessNear-lossless for most tasks | ≈2.1 GB |
| Q6_KNear-losslessPractically lossless | ≈2.3 GB |
| Q8_0LosslessLossless in practice, twice the size of Q4 | ≈2.7 GB |
| FP16 / BF16LosslessFull precision, as released | ≈4.3 GB |
Estimates for 8K tokens of context: weights + KV cache + runtime overhead. Speed assumes memory-bound decoding at your GPU’s bandwidth. Pick a GPU above and the fit column checks every row against it. Weights marked “file” use the real GGUF download size; “est.” rows are computed from the parameter count.
How to pick a quant
Take the biggest quant that fits with room left for context. Q4_K_M is the usual sweet spot. Below Q4 quality drops fast, above Q6 you mostly pay memory for nothing. Big models (70B+, large MoE) tolerate low quants better than small ones.
- LosslessF16, Q8_0
- Near-losslessQ6_K, Q5_K_M
- Small lossQ5_K_S, Q4_K_M, IQ4_NL, IQ4_XS, Q4_K_S
- Noticeable lossQ4_0, Q3_K_L, Q3_K_M, IQ3_M
- Heavy lossIQ3_XXS, Q3_K_S, IQ2_M, Q2_K
Typical quality loss, not measured per model. Tiers come from KL divergence against FP16 on Llama 3 8B: llama.cpp perplexity README: LLaMA 3 8B scoreboard (KLD and ΔPPL vs FP16, Wikitext-2); llama.cpp quantize tool: ΔPPL per quant type @ Llama-3-8B
How much context fits?
At 8K context, Q4_K_M needs ≈1.9 GB(1.0 GB weights + 0.4 GB KV cache + 0.6 GB overhead). Pick a GPU above to see how much context fits.
Weights + overheadKV cache
Can the RTX 4060 run LFM3-VL-2B? →
Speed vs quality per quant
Smaller quants run faster and fit in less memory, but lose quality.
Fits in VRAMTight fitCPU offload✕ won't runQuality is typical KL divergence vs FP16, not measured for this model.
Real-world speed of LFM3-VL-2B
Tokens per second measured by people running it at home. A GPU shows a median once 3 reports agree on the same setup. Reports far from our estimate wait for a review. Measure yours with llama-bench →
No medians yet. Be one of the first 3 to report a setup.
All reports
Run LFM3-VL-2B locally
--quantQ4_K_Mdefault; pick your GPU to tailor itLiquidAI/LFM3-VL-2B-GGUF
ollama run lfm3-vl:2bThe library build picks its own default quantisation.
ollama run hf.co/LiquidAI/LFM3-VL-2B-GGUF:Q4_K_M llama-server -hf LiquidAI/LFM3-VL-2B-GGUF:Q4_K_MDownloads the GGUF on first run. Installed through llama.app? The same flags work with `llama serve`.
lms get LiquidAI/LFM3-VL-2B-GGUF@Q4_K_MThen load it from the LM Studio app or with `lms load`.
from transformers import pipeline
pipe = pipeline("image-text-to-text", model="LiquidAI/LFM3-VL-2B", dtype="auto", device_map="auto")
messages = [
{
"role": "user",
"content": [
{"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
{"type": "text", "text": "Describe this image."},
],
}
]
out = pipe(text=messages, max_new_tokens=256)
print(out[0]["generated_text"][-1]["content"])Quantised versions and downloads
| Repository | Format | Quant | Size |
|---|---|---|---|
| LiquidAI/LFM3-VL-2B-NVFP4 | Safetensors | NVFP4 | — |
| LiquidAI/LFM3-VL-2B-FP8 | Safetensors | FP8 | — |
| bartowski/LFM3-VL-2B-GGUF | GGUF | IQ2_M | 0.7 GB |
| bartowski/LFM3-VL-2B-GGUF | GGUF | Q2_K | 0.7 GB |
| bartowski/LFM3-VL-2B-GGUF | GGUF | IQ3_XXS | 0.8 GB |
| bartowski/LFM3-VL-2B-GGUF | GGUF | Q3_K_S | 0.8 GB |
| bartowski/LFM3-VL-2B-GGUF | GGUF | IQ3_XS | 0.8 GB |
| bartowski/LFM3-VL-2B-GGUF | GGUF | Q3_K_M | 0.8 GB |
| bartowski/LFM3-VL-2B-GGUF | GGUF | Q3_K_L | 0.9 GB |
| bartowski/LFM3-VL-2B-GGUF | GGUF | IQ3_M | 0.9 GB |
About LFM3-VL-2B
LFM3-VL-2B is a vision-language model for on-device visual understanding. It takes text, images, and video as input, and combines the LFM3 language model with a new in-house vision encoder. New in LFM3-VL-2B: • Video understanding: the first LFM vision model that takes video as input. Ask questions about a recorded lecture, a walkthrough filmed on your phone, or a clip from a camera feed. • Thinking: the model reasons step by step before it answers. This helps most with harder visual questions, such as math, charts, and multi-step problems. • Spatial and embodied reasoning: the model finds objects and reasons about how they are placed relative to each other, a strong foundation for robotics applications. Find more information about the LFM3 family in our release blog post. The LFM3 family includes base, text, vision, audio, omni, and decision models.
Full model card on Hugging Face →Alternatives to LFM3-VL-2B
All alternatives →MiniCPM-V 4.7 is the latest model in the MiniCPM-V series, available in a 1B variant. It builds on the family's single-image, multi-image, and video understanding capabilities, combining stronger foundation capabilities with efficient visual encoding and on-device inference.
Meet Qwen3-VL — the most powerful vision-language model in the Qwen series to date.
d1-2B is a decision model for text, pictures, and video, built on LFM3-VL-2B. It returns calibrated answers in one forward pass, without generating tokens.
Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance.
Muse Glimmer is a 30-billion-parameter causal language model with a dedicated perception encoder, distilled from Muse Spark and purpose-built for autonomous agentic tasks on consumer hardware.
LFM3-VL-2B FAQ
- How much VRAM does LFM3-VL-2B need?
- About 2.0 GB at Q4_K_M with an 8K context. Q8_0 needs about 2.9 GB and full precision about 4.6 GB. Longer contexts need more memory for the KV cache.
- Can LFM3-VL-2B run on the RTX 4060?
- Yes. It fits at FP16 / BF16 (about 4.3 GB), at an estimated 39 tokens/s. It fits comfortably on the Apple M4 (16 GB), Radeon RX 6700 XT, RTX 3060 12GB.
- Can I use LFM3-VL-2B commercially?
- Yes, with conditions. It is released under lfm1.0. Custom licence: read the model card before using it commercially.
- What is the context length of LFM3-VL-2B?
- 128,000 tokens (125K). The full window needs far more memory than an 8K context.
- How do I run LFM3-VL-2B with Ollama or llama.cpp?
- With Ollama: ollama run lfm3-vl:2b. With llama.cpp: llama-server -hf LiquidAI/LFM3-VL-2B-GGUF:Q4_K_M, which starts an OpenAI-compatible server.
Rate this model
Data from Hugging Face, the model card and public leaderboards. Memory and speed figures are estimates; real numbers depend on the runtime, context length and drivers. Always check the licence yourself before commercial use.
0 comments