$ cat wiki/models/lfm2-5-24b-a2b.md
LFM2.5-24B-A2B
Compared with
- Beam 1.0
- EmbeddingGemma 2
- Mistral Large 4
- Kolibri-1
- GPT-6.1 Sol
- MiMo-V2.6-Pro
- Grok 4.7
- Ternary Bonsai 2 27B
- Fugu Max
- Kimi K2.8 Preview
- DeepSeek V4.1-Flash
- K2 Horizon
- Hy4 preview
- GLM-5.3-Flash
- Granite 4.2
- Ling-3.0-tiny
- Laguna S 2.1
- Inkling
- LongCat-2.0
- Gemini 4 Argon
- Claude Sonnet 5.5
- Claude Opus 5.5
- GPT-6 Luna
- GPT-6 Sol
- Gemini 3.8 Live
- Fugu Ultra v2
- GPT-Image-2.5 Flare
- GPT-Image-2.5 Sunburst
- Astra
- Gemini 3.8 Flash
- Muse Spark 1.3
- Claude Fable 5.1
- GLM-5.3
- Qwen 3.8 27B
- DeepSeek V4-Pro-0813
- Gemini 3.7 Flash
- Muse Glimmer
- Grok 4.6
- Muse Spark 1.2
- Qwen 3.8 Max
- DeepSeek V4-Flash
- Claude Opus 5
- Gemini 3.5 Flash-Lite
- Gemini 3.6 Flash
- DeepSeek V4
- Kimi K3
- GPT-5.6 Sol
- Grok 4.5
- Claude Sonnet 5
- GLM-5.2
- Claude Fable 5
- MiniMax M3
- Claude Opus 4.8
- Gemini 3.5 Flash
- Grok Build
- Claude Opus 4.7
- Muse Spark
Spec
| Attribute | Value |
|---|---|
| Developer | Liquid AI |
| Released | 2026-10-13 |
| Announced | 2026-10-13 |
| Context window | 131,072 tokens |
| Pricing | $0.10/M input · $0.05/M cached input · $1.00/M output |
| Catalogue pricing | $0.05/M input · $0.70/M output (Liquid FP8 hosted endpoint; OpenRouter, 2026-10-22) |
| License | unknown |
| Availability | Downloadable base and post-trained weights; Liquid Playground and API |
| The launch describes a 24B-parameter model with 2B active parameters, 40 layers, 64 experts per MoE block and top-4 routing. It links a model license, whose terms were not inspected; open weights do not establish an unrestricted license. (source) |
Release Date
Liquid AI released the base and post-trained models on 2026-10-13. The post-trained model always emits reasoning before its answer. (source)
Benchmarks
Liquid AI reports LiveClaw 48.09 versus 38.19 for Qwen3.6-35B-A3B and AutomationBench 27.69 versus 8.72, but SWE-Bench Verified with OpenCode 51.00 versus 66.20. These are vendor measurements with benchmark-specific generation settings, not universal superiority. (source)
The reported 184 tokens/s on Apple M5 Max measures device decoding. The separate over 21K output tokens/s on one NVIDIA B300 is aggregate throughput under sustained high concurrency, SGLang 0.5.22, BF16, 1,024 input tokens and up to 256 output tokens. The two throughput figures describe different workloads. (source)
Use Cases
Local agents and tool use are the intended workloads. Post-training combines SFT, teacher specialization, non-agentic and agentic on-policy distillation, then agentic RL inside execution harnesses. Liquid AI reports discovering reward hacks during teacher specialization, tightening sandboxes, and using sealed environments for the published evaluations. (source)
Compared To
The model trades stronger tool-use results against weaker knowledge and coding results in the vendor's comparison. Its training and evaluation design connects to Agentic Reinforcement Learning and Eval Harness Configuration; workload-specific testing remains necessary. (source)
Conflicting Reports
On 2026-10-22 the Liquid announcement still states $0.10/M input, $0.05/M cached input and $1.00/M output for its API. The Liquid FP8 offer on OpenRouter instead quotes $0.05/M input and $0.70/M output. These are different serving offers; the vendor API figure is retained, and the catalogue row is checked against the hosted offer only.