LiquidAI’s LFM3: Six Specialized Models, Benchmark Claims, and Evaluation Gaps

LiquidAI presents six LFM3 variants—text, vision, audio, omni, and two decision models—built from the LFM3-1.7B-Base backbone. The launch post reports gains on selected benchmarks, but those results do not establish how a model will perform on a particular production workload. ¹

What LiquidAI says it shipped with LFM3

The six models are LFM3-1.7B, LFM3-VL-2B, LFM3-Audio-2B, LFM3-Omni-2B, d1-2B, and d1-omni-2B. LiquidAI describes d1-2B as providing “calibrated answers,” but the post excerpt does not explain the decision models’ intended tasks or evaluation. ¹

The practical question is whether a narrower model improves performance or deployment efficiency on a defined workflow enough to justify evaluating and operating another model path. Specialization can mean using a model for inputs it is designed to handle; it can also mean maintaining and validating more variants. The launch post describes a model family and reports selected results, but those results do not establish performance on your workload. ¹

LFM3 model variants and the inputs they handle

Variant	Listed modality or role	What remains unclear from the post excerpt
LFM3-1.7B	Text; LiquidAI recommends it for local agents, tool calling, structured outputs, or as a fine-tuning base. 1	Performance and deployment requirements for a specific workload.
LFM3-VL-2B	Text, images, and video. 1	Video limits, supported formats, and performance on a team’s visual tasks.
LFM3-Audio-2B	Text and audio, including speech. 1	Audio limits, supported formats, and task-specific behavior.
LFM3-Omni-2B	Text, images, video, and audio. 1	Input boundaries and resource requirements.
d1-2B	Decision model; described as producing calibrated answers. 1	Intended tasks and evaluation details.
d1-omni-2B	Decision model. 1	Modalities and intended tasks.

LiquidAI presents the variants as options for different modalities and says they can be deployed where an application runs. The post excerpt does not specify video or audio limits, supported formats, or exact output behavior for each model. ¹ Input compatibility can help shortlist candidates among multimodal AI models; by itself, it says little about accuracy, latency, robustness, or failure behavior on your data.

LiquidAI LFM3 benchmark results

Comparison	Reported result
LFM3-1.7B vs. LFM2.5-2.6B	LFM3-1.7B scores 38.42 on WebArena-Verified versus 3.02 for LFM2.5-2.6B. LiquidAI says LFM3-1.7B uses 35% fewer parameters. 1
LFM3-VL-2B vs. LFM2.5-VL-3B	LiquidAI reports 1.9B versus 3.1B parameters and an average of 71.4 versus 57.6 across 62 vision benchmarks. It says the result is on par with Gemma 4 12B. 1
LFM3 vision encoder vs. Qwen3.5-2B vision encoder	LiquidAI says its 164M-parameter encoder matches accuracy with 2.2 times fewer FLOPs, measured using the same LFM2.5-230M language backbone. 1

These figures are benchmark results reported by LiquidAI. They can inform a shortlist, but they do not by themselves establish performance on a different task or in an application’s full production pipeline. Teams should test their own quality, operational, and risk requirements rather than treating benchmark scores as workload forecasts. ¹

What specialized AI models may offer

LiquidAI says the text model uses about 1 GB at 4-bit precision and has a 256K-token context window. It positions LFM3-1.7B for on-device agents. ¹ The post also says LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices and is about 25% faster with 8K tokens in context. ¹

For audio, LiquidAI says its causal encoder can prefill while a person is speaking, with time to first token as low as 50 ms. It reports that LiquidCodec generates speech at about 87 times real time on a single CPU thread on an Apple M4 Max and supports streaming decoding. ¹

Treat these as scenario-specific figures, not application-level forecasts. A model-level result does not establish how a system will perform after preprocessing, retrieval, tool calls, generation, post-processing, and monitoring. A specialized model earns a place when its measured advantage holds in the full pipeline and is worth the cost of maintaining a separate path.

If you are considering LFM3-1.7B for tool calling, the test set matters as much as the model. Tool-Using Agents Need Tests for Users Who Leave Things Unsaid covers why clean, fully specified requests flatter an agent.

Fine-tuning LFM3 on your own data: a starting point, not a plan

LiquidAI says each model can be fine-tuned on a single GPU using a team’s own data. ¹ The post excerpt does not specify a GPU model, memory requirement, training duration, dataset size, fine-tuning methods, or model-specific procedure. ¹

Before using “single GPU” as a planning assumption, ask LiquidAI or consult the relevant model card for:

• GPU model and memory requirement
• Fine-tuning method, settings, and expected duration
• Data format, dataset size, and preparation requirements
• Checkpoint handling and validation procedure

The training run is only one part of the work, and custom multimodal models add data-preparation questions for each input type. A useful pilot compares the base model, a fine-tuned version, and a suitable alternative, drawn from a governed model catalog or elsewhere, on the same data and conditions. Measure task quality, latency, memory, cost, and failure modes; include held-out examples and inputs outside the training distribution.

LFM3 deployment details to verify

LiquidAI says the LFM3 models run in Transformers and that the model cards list other supported inference engines. ¹ The post excerpt does not name those engines or provide hardware and deployment constraints for each variant.

The text model’s memory and context figures are starting references, not deployment guarantees. The post excerpt gives no equivalent memory or context figures for the other variants. ¹ Before committing, test the model on the intended device and backend with realistic inputs. Track peak memory, sustained throughput, time to first token, total response latency, power draw, and concurrency under your expected workload.

If a pilot ends with several model paths in use, governance becomes its own workload. Yelno’s enterprise AI gateway offers routing constraints for region, data policy, and allowed providers and models, plus budgets and audit logs. The LFM3 post does not say whether these models are available through any gateway, so check availability separately.