Liquid AIAudio modelOpen weights
LFM3-Audio-2B
Summary
LFM3-Audio-2B accepts text and 16 kHz audio and returns text, 24 kHz speech, or interleaved text and speech. Its causal audio encoder and LiquidCodec support streaming. The model has about two billion parameters and 32K context tokens.
The publisher card lists ASR, TTS, VoiceBench, internal tool calling, and MMAR separately. WER and CER are error rates; UTMOSv2 uses a 1 to 5 scale. Publisher claims about codec real-time speed and first-token latency are kept apart from the benchmark table and are not general runtime guarantees.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 8 published benchmark results
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
Multilingual ASR WER / CER (Show measurement, test conditions, and source)
- Source value
- 6.03
- Score
- 6.03
- Metric
- WER / CER
- Unit
- %
- Category
- Audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- LiquidAI/LFM3-Audio-2B official model card at revision cddb56b5ec9b4703f9b3b2d0500143d1fff2fee0
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Liquid AI model-card evaluation; dataset-specific procedures are stated in the model card. ASR is an error rate, TTS uses UTMOSv2, and the internal tool-call score is a percentage.
- Context
- Manufacturer-reported; these sources do not provide an independent reproduction of this result.
VoiceBench (Show measurement, test conditions, and source)
- Source value
- 73.13
- Score
- 73.13
- Metric
- percent
- Unit
- %
- Category
- Audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- LiquidAI/LFM3-Audio-2B official model card at revision cddb56b5ec9b4703f9b3b2d0500143d1fff2fee0
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Liquid AI model-card evaluation; dataset-specific procedures are stated in the model card. ASR is an error rate, TTS uses UTMOSv2, and the internal tool-call score is a percentage.
- Context
- Manufacturer-reported; these sources do not provide an independent reproduction of this result.
Tool calling internal eval (Show measurement, test conditions, and source)
- Source value
- 58.5
- Score
- 58.5
- Metric
- percent
- Unit
- %
- Category
- Audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- LiquidAI/LFM3-Audio-2B official model card at revision cddb56b5ec9b4703f9b3b2d0500143d1fff2fee0
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Liquid AI model-card evaluation; dataset-specific procedures are stated in the model card. ASR is an error rate, TTS uses UTMOSv2, and the internal tool-call score is a percentage.
- Context
- Manufacturer-reported; these sources do not provide an independent reproduction of this result.
Audio understanding MMAR (Show measurement, test conditions, and source)
- Source value
- 50.6
- Score
- 50.6
- Metric
- percent
- Unit
- %
- Category
- Audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- LiquidAI/LFM3-Audio-2B official model card at revision cddb56b5ec9b4703f9b3b2d0500143d1fff2fee0
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Liquid AI model-card evaluation; dataset-specific procedures are stated in the model card. ASR is an error rate, TTS uses UTMOSv2, and the internal tool-call score is a percentage.
- Context
- Manufacturer-reported; these sources do not provide an independent reproduction of this result.
English ASR WER (Show measurement, test conditions, and source)
- Source value
- 6.23
- Score
- 6.23
- Metric
- WER
- Unit
- %
- Category
- Audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- LiquidAI/LFM3-Audio-2B official model card at revision cddb56b5ec9b4703f9b3b2d0500143d1fff2fee0
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Liquid AI model-card evaluation; dataset-specific procedures are stated in the model card. ASR is an error rate, TTS uses UTMOSv2, and the internal tool-call score is a percentage.
- Context
- Manufacturer-reported; these sources do not provide an independent reproduction of this result.
English TTS UTMOSv2 (Show measurement, test conditions, and source)
- Source value
- 3.83
- Score
- 3.83
- Metric
- UTMOSv2 on a 1 to 5 scale
- Unit
- rating
- Category
- Audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- LiquidAI/LFM3-Audio-2B official model card at revision cddb56b5ec9b4703f9b3b2d0500143d1fff2fee0
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. Liquid AI model-card evaluation; dataset-specific procedures are stated in the model card. ASR is an error rate, TTS uses UTMOSv2, and the internal tool-call score is a percentage.
- Context
- Manufacturer-reported; these sources do not provide an independent reproduction of this result.
Audio time to first token, reported best case (Show measurement, test conditions, and source)
- Source value
- 50
- Score
- 50
- Metric
- time to first token upper bound
- Unit
- ms
- Category
- Streaming latency
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Liquid AI LFM3 release announcement
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. LFM3-Audio causal audio encoder streams while the user speaks; Liquid AI says time to first token can be as low as 50 ms. The source does not state hardware, input clip, workload or exact checkpoint measurement procedure.
- Context
- Manufacturer-reported best case. 50 ms is the stated low-end latency, not an SLA or general measurement.
LiquidCodec speech decoding speed on Apple M4 Max (Show measurement, test conditions, and source)
- Source value
- 87
- Score
- 87
- Metric
- real-time factor
- Unit
- ratio
- Category
- Codec throughput
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Liquid AI LFM3 release announcement
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. LiquidCodec generated-speech decoding on a single Apple M4 Max CPU thread; manufacturer reports about 87x real time. The claim concerns the codec, not end-to-end LFM3-Audio model inference.
- Context
- Manufacturer-reported codec result. This is not a model benchmark or end-to-end latency.
All specifications
Architecture
- Architecture
- LFM3 backbone with a causal audio encoder and LiquidCodec audio tokenizer. Supports streaming and incremental prefill.
- Source: Architecture
Input and output
Licence and access
- Access
- Open weights, local deployment, and the API and Playground linked by the publisher
- Source: Access
Further details
- Speech languages and locales
- 18 listed speech languages and locales, including British and American English.
- Source: Speech languages and locales
- Speech generation
- The model card documents streaming, audio_temperature=1.0, audio_top_k=128, and an audio_repetition_window of 4,096 for the speech recipe.
- Source: Speech generation
Sources and data date
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Without vendor claims and historical or disputed results; those stay labelled in the single values.
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 29, 2026. 62 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | ASR sample counts Note The card names the ESB/Open-ASR suite and nine FLEURS languages but gives no per-language audio-file counts. |
| Not published | Internal tool-call task set Note The internal tool-calling evaluation publishes no task count, speakers, tool schemas, or error definition. |
| Not found in the inspected sources | Independent evaluation Note The reviewed manufacturer sources do not link an independent reproduction of the LFM3-Audio results. |
| Not found in the inspected sources | Technical report or research paper Note Reviewed publisher sources do not link a separate technical report or research paper for LFM3-Audio-2B. |
| Not found in the inspected sources | Vals rows in captured payloads Note Nine complete Vals payloads dated 28 October were checked; no exact model ID or unambiguous model label. Limited to these saved benchmarks. |
| Unresolved | Uncaptured Arena Code/Speech leaderboards Note Text, Agent including Code/Chat/Work, and twelve WebDev views were captured. A separate general Code and Arena Speech leaderboard were not saved as complete tables. |
| Unresolved | Artificial Analysis catalog reconciliation Note The full current AA model and benchmark catalog was not re-reconciled for this profile in this sidecar. This sidecar makes no AA absence claim; native AA imports remain authoritative. |
| Not found in the inspected sources | VoiceArena STT exact profile row Note The saved VoiceArena speech-to-text results do not contain this exact model identity. This is limited to that capture and does not rule out future rows or other speech evaluators. |
| Additional source | LiquidAI/LFM3-Audio-2B model card at pinned revision cddb56b5ec9b4703f9b3b2d0500143d1fff2fee0 (retrieved October 29, 2026) · LFM3-Audio-2B model card · Editorial description reviewed October 29, 2026 |
| Additional source | Liquid AI LFM3 release announcement (retrieved October 29, 2026) · Liquid AI LFM3 release announcement · Editorial description reviewed October 29, 2026 |
| Additional source | Liquid AI model documentation (retrieved October 29, 2026) · Liquid AI audio documentation · Editorial description reviewed October 29, 2026 |
| Additional source | Liquid AI LFM3 Hugging Face collection (retrieved October 29, 2026) |
| Additional source | Hugging Face model metadata (retrieved October 29, 2026) |
| Additional source | Model metadata (retrieved October 29, 2026)123 |