AI Launch Profile
Liquid AI LFM2.5-2.6B
Liquid AI released the base and post-trained LFM2.5-2.6B checkpoints on August 4, 2026. The post-trained text model targets agentic workloads and ships in native, GGUF, MLX, and ONNX formats, with support for Transformers, llama.cpp, vLLM, SGLang, and other local inference tools.
At a glance
Launch Snapshot
- Company
- Liquid AI
- Launch date
- August 4, 2026
- Launch type
- Not classified
- Category
- AI Agents
- Audience
- Developers, AI engineers, edge-device teams, privacy-sensitive organizations, and local-AI users who can evaluate model quality and operate their own inference runtime
- Pricing
- The model weights are downloadable under Liquid AI's LFM1.0 license. Liquid AI did not announce a paid hosted API price for this release, and the reviewed Hugging Face page says no inference provider currently deploys the model. Self-hosting avoids a vendor per-token fee but still carries hardware, electricity, engineering, and support costs.
- Free plan
- Not publicly confirmed
- API
- Not publicly confirmed
- Open weights/source
- Yes
Launch Context
Use these links to move from this record into the broader Launch Intelligence database.
Verification & Sources
- Status
- Verified
- Source links
- 5
- Freshness
- Verified August 7, 2026
- Last verified
- August 7, 2026
- Last updated
- August 7, 2026
Key source checks
Suggest a correction
Kingy Scores
Scores are editorial review signals across launch quality, demo evidence, YouTube potential, and search readiness.
- Launch Score
- 9.2 / 10
- Demo Quality
- 7.5 / 10
- YouTube Potential
- Medium
Creator Coverage Next Steps
This launch has signals that may support demos, reviews, creator education, founder storytelling, or practical product explainers.
Launching an AI product that needs clear demos, creator education, and buyer trust? Sponsor a Kingy AI video or launch feature.
Kingy AI Take
LFM2.5-2.6B is a credible option for high-volume, privacy-sensitive local agents where compact size and tool use matter more than frontier reasoning. Treat it as a workflow component to benchmark on the exact task, not as a small replacement for a strong cloud model.
Who it is for
Developers, AI engineers, edge-device teams, privacy-sensitive organizations, and local-AI users who can evaluate model quality and operate their own inference runtime
What feels promising
Day-one native, GGUF, MLX, and ONNX releases make the model unusually easy to test across laptops, CPUs, GPUs, and edge devices. The 128K context window and native tool-call format are directly relevant to long-running agent traces.
What feels unproven
The benchmark and speed figures in the reviewed launch materials come from Liquid AI and were not independently reproduced for this record. The model card says the model is not recommended for agentic coding or knowledge-heavy tasks. A 128K advertised context window does not prove reliable reasoning across the full window. LFM1.0 is a model license, so teams should review its terms rather than assume an OSI open-source software license. Mobile and desktop throughput will vary with quantization, memory bandwidth, prompt length, runtime, and thermals.
Traction notes
The official Hugging Face model page showed 77,973 downloads in the prior month at review time. That is a platform activity signal, not proof of production adoption, retention, or business traction.
Source-backed record