Data as of October 27, 2026 · How the score is built
Released Oct 27, 2026128K contextLiquid AI LFM3-VL-2B model card
LFM3-VL-2B
Decision readingLFM3-VL-2B is tracked, but not publicly ranked yet. The profile exposes 20 sourced benchmark rows and leaves unsupported fields blank until a published record exists.
Released Oct 27, 2026 — see all recent releases
Decision snapshot
Each value carries a field reference instead of floating alone. Markers compare this model with the current ranked and priced catalog; they are not absolute quality thresholds.
Capability
Unranked
field median 50.2Not eligible for a public rank
Price
Self-hosted; infrastructure cost varies
input median $0.75No comparable first-party hosted token rate
Speed
Not measured
field median 96 tok/sTime to first token not measured
Context
128Ktokens
field median 256,000Maximum output length is tracked separately
Strongest published evidence
Published rows are visible, but no category has enough eligible evidence for a comparative rank.
Validate before choosing
20 published rows leave some tracked benchmark slots empty. No comparable first-party API token rate is published.
Source-linked · 20 displayable benchmark rows
Category score record
Scores and ranks appear only where published evidence can be displayed. The table keeps the score, weight, cohort, and evidence state together.
| Category | Score | Rank | Percentile | Weight | Benchmarks | Evidence |
|---|---|---|---|---|---|---|
| AgenticWeight 22%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| CodingWeight 20%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| ReasoningWeight 17%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| MultimodalRank Not rankedWeight 12%20 benchmarksVerified | 19.6 | 20 benchmarks | Verified | |||
| KnowledgeWeight 12%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| MultilingualWeight 7%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| Inst. FollowingWeight 5%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| MathWeight 5%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured |
20 of 774 tracked benchmark slots have displayable evidence · bars run 0–100
Coverage detailsHow much of this is verified
Coverage is split by category so a strong number never hides a thin evidence base. Verified means the row is tied to a published source; provisional rows remain visible but separate.
- AgenticNot measured
- CodingNot measured
- ReasoningNot measured
- Multimodal20/20 verified
- KnowledgeNot measured
- MultilingualNot measured
- Inst. FollowingNot measured
- MathNot measured
Benchmark ledger
Multimodal opens by default. The marker compares each value with the best source-verified result in the catalog; provisional leaders do not set the reference. Expand the remaining categories for every published row.
Multimodal20 rows
| Benchmark | Score | Versus best verified row | Gap | Weight | Evidence |
|---|---|---|---|---|---|
| MMMU-ProMassive Multi-discipline Multimodal Understanding Pro | Score45.7% | Versus best verified row Best verified: GPT-6.1 Terra · 85.5% | Gap39.8 behind | WeightWeighted 40% | Provider exact |
| RealWorldQA | Score79.3% | Versus best verified row Best verified: Qwen3.8-Flash-Next · 88.5% | Gap9.2 behind | WeightDisplay only | Provider exact |
| SimpleVQA | Score36.3% | Versus best verified row Best verified: Qwen3.7 Plus · 81.7% | Gap45.4 behind | WeightDisplay only | Provider exact |
| CountBench | Score93.0% | Versus best verified row Best verified: Qwen3.6-27B · 97.8% | Gap4.8 behind | WeightDisplay only | Provider exact |
| MMMUMassive Multi-discipline Multimodal Understanding | Score55.7% | Versus best verified row Best verified: Qwen3.5 397B · 85.0% | Gap29.3 behind | WeightDisplay only | Provider exact |
| OCRBench V2 | Score63.5% | Versus best verified row Best verified: Qwen3.8 Max · 74.2% | Gap10.7 behind | WeightDisplay only | Provider exact |
| RefCOCO (avg)RefCOCO average | Score90.5% | Versus best verified row Best verified: Qwen3.6-27B · 92.5% | Gap2 behind | WeightDisplay only | Provider exact |
| V* | Score86.4% | Versus best verified row Best verified: Qwen3.6 Plus · 96.9% | Gap10.5 behind | WeightDisplay only | Provider exact |
| CharXiv w/o toolsCharXiv Reasoning without tools | Score70.2% | Versus best verified row Best verified: Kimi K3.1 Pro · 89% | Gap18.8 behind | WeightDisplay only | Provider exact |
| MathVision | Score51.6% | Versus best verified row Best verified: Qwen3.8 Max · 95.2% | Gap43.6 behind | WeightDisplay only | Provider exact |
| DynaMath | Score79.2% | Versus best verified row Best verified: Qwen3.5 397B · 86.3% | Gap7.1 behind | WeightDisplay only | Provider exact |
| We-Math | Score68.6% | Versus best verified row Best verified: Qwen3.5 397B · 87.9% | Gap19.3 behind | WeightDisplay only | Provider exact |
| ERQA | Score48.2% | Versus best verified row Best verified: Qwen3.8 Max · 77.8% | Gap29.6 behind | WeightDisplay only | Provider exact |
| ScreenSpot Pro | Score55.5% | Versus best verified row Best verified: GPT-6 Astra · 92.7% | Gap37.2 behind | WeightDisplay only | Provider exact |
| LVBench | Score41.2% | Versus best verified row Best verified: Gemini 4 Argon · 91.7% | Gap50.5 behind | WeightDisplay only | Provider exact |
| MLVU (M-Avg)MLVU mean average | Score71.3% | Versus best verified row Best verified: Qwen3.8 Max · 90.8% | Gap19.5 behind | WeightDisplay only | Provider exact |
| MMVUMultimodal Multi-disciplinary Video Understanding | Score52.9% | Versus best verified row Best verified: Qwen3.8 Max · 82.4% | Gap29.5 behind | WeightDisplay only | Provider exact |
| Video-MME (w/o subtitle)Video-MME without subtitle | Score62.8% | Versus best verified row Best verified: Qwen3.5 397B · 83.7% | Gap20.9 behind | WeightDisplay only | Provider exact |
| Video-MME (with subtitle)Video-MME with subtitle | Score69.0% | Versus best verified row Best verified: Qwen3.8 Max · 90.4% | Gap21.4 behind | WeightDisplay only | Provider exact |
| VideoMMMU | Score49.6% | Versus best verified row Best verified: Qwen3.8 Max · 88.7% | Gap39.1 behind | WeightDisplay only | Provider exact |
Bars run 0–100; the dark tick marks the best source-verified value
All 20 rowsLineage
The sequence follows explicit supersedes links. Each score is estimated for that model; a relative can inform a sparse estimate but never sets a floor, so a newer release can score below an earlier one. Scores and prices remain blank when the corresponding public row or first-party rate is unavailable.
Reasoning
Spec sheet
Each documented value carries its source. Missing fields stay visible as not sourced or not published, rather than disappearing from the page.
- API model ID
- LiquidAI/LFM3-VL-2BLiquid AI LFM3-VL-2B model card
- Context window
- 128KLiquid AI LFM3-VL-2B model card
- Maximum output
- Not sourced yet
- Knowledge cutoff
- Not sourced yet
- Input modalities
- Not sourced yet
- Output modalities
- Not sourced yet
- Parameters
- Not sourced yet
- Availability
- Liquid AI publishes the 1.92B-parameter vision-language model (1.73B LFM3 language model, 164M in-house vision encoder, 25M projector) under the LFM Open License v1.0, with BF16, FP8, NVFP4, GGUF, MLX and ONNX checkpoints and a 131,072-token context. It takes text, images and video (sampled at 2 fps) and thinks before answering. At launch, llama.cpp and SGLang support are still in review and vLLM support is in progress.
- Cloud regions
- Not tracked yet
- Lifecycle
- Current
- API capabilities
- Tool calling, structured outputs, and batch support are not tracked yet
- Prompt caching
- Not documented in the pricing recordLiquid AI LFM3-VL-2B model card
- Self-host
- LFM3-VL-2B needs ~24GB VRAM (1× NVIDIA RTX 4090 (24GB)).
- Rate limits
- Not tracked yet
How to read this profile
The visual layer above carries the decisions. These notes preserve the model, ranking, coverage, and family context behind the numbers.
We track LFM3-VL-2B, but the public leaderboard excludes this profile until enough non-generated benchmark coverage is available. Only published rows appear above.
LFM3-VL-2B is a open weight model with a 128K context window. It uses an explicit reasoning mode, which can improve complex problem solving while adding latency and token use.
Liquid AI publishes the 1.92B-parameter vision-language model (1.73B LFM3 language model, 164M in-house vision encoder, 25M projector) under the LFM Open License v1.0, with BF16, FP8, NVFP4, GGUF, MLX and ONNX checkpoints and a 131,072-token context. It takes text, images and video (sampled at 2 fps) and thinks before answering. At launch, llama.cpp and SGLang support are still in review and vLLM support is in progress.
Official exact-value snapshot from the LFM3-VL-2B Hugging Face model card (reasoning-mode column) and the October 27, 2026 LFM3 launch post. BenchLM maps RealWorldQA, SimpleVQA, CountBenchQA, MMMU (val), MMMU-Pro (10-option multiple choice), OCRBench v2 (English), RefCOCO average, V*Bench, CharXiv reasoning (val, no tools), MATH-Vision (full), DynaMath average accuracy, We-Math (testmini, strict), ERQA, ScreenSpot-Pro, LVBench, MLVU (dev, M-Avg), MMVU (val), Video-MME without and with subtitles, and Video-MMMU onto existing lanes. The other 41 of Liquid's 62 vision benchmarks, including MMStar, MME, MMBench, DocVQA, ChartQA, POPE, HallusionBench and ScreenSpot-v2, have no exact compatible lane. Liquid ran every model with vLLM 0.27.1; all values are provider-run.
20 of 774 tracked benchmark slots currently have displayable evidence. Missing categories stay blank.
Last updated October 27, 2026. Runtime fields remain blank until a sourced snapshot exists.
Deployment options
Self-host and provider-specific paths stay separate from benchmark evidence so operating constraints are visible before a score becomes the whole decision.
Self-host vs API cost
Estimates at 50,000 req/day · 1000 tokens/req average.
Questions
How does LFM3-VL-2B perform overall in AI benchmarks?
LFM3-VL-2B has 20 source-displayable benchmark rows, but it does not qualify for a public overall rank. The available rows remain visible by category without being converted into a site-wide score. Missing evidence stays blank instead of being estimated from an earlier model.
Is LFM3-VL-2B good for multimodal and grounded tasks?
LFM3-VL-2B has source-displayable benchmark coverage for multimodal and grounded tasks, but the public category table does not assign it a rank there. The individual rows remain available for inspection. A missing category position means the evidence threshold was not met; it does not convert the model's unmeasured work into a zero.
Is LFM3-VL-2B open source?
LFM3-VL-2B is an open-weight model from LiquidAI. Its weights can be downloaded for local or hosted deployment, subject to the published license. Open weight does not automatically mean open source: training data and training code may remain private, and commercial restrictions can still apply.
Does LFM3-VL-2B have full benchmark coverage on BenchLM?
No. LFM3-VL-2B currently has 20 source-displayable rows. The profile exposes published, non-generated evidence and leaves missing categories blank until an exact evaluation is available. Coverage describes how much was measured; it is not a penalty added to an individual benchmark result.
What is the context window size of LFM3-VL-2B?
LFM3-VL-2B has a documented context window of 128K. That figure is the maximum combined prompt and retained-conversation space reported for this exact model; it is not the maximum output length. The profile keeps output limits separate because providers often publish those limits independently.