AI is set to reshape workplace hierarchies by giving individuals the power to do work that once required entire teams.
Liquid AI, founded by former MIT computer scientists, today released its smallest AI language model yet, LFM2.5-230M, and enterprises would do well to consider it for their uses in data extraction and local deployment on smartphones, laptops and robotics.
This is a 230-million-parameter foundation model explicitly designed for on-device agentic workflows, and as Liquid states in its release blog post, that small size makes it possible to run nearly “anywhere.” According to Liquid, it also outperforms models more than 4X its size on selected benchmarks, specifically doing better at data extraction than the 800 million parameter count Alibaba Qwen3.5-0.8B (Instruct) and 1-billion parameter Google Gemma 3 1B.
The model targets developers and engineers building lightweight data extraction pipelines and autonomous edge systems.
Operating under a dual-use commercial license, the model remains free for individuals and companies generating less than $10 million in annual revenue, while requiring a paid enterprise agreement for larger corporations.
This release distinguishes itself from other small AI models by utilizing the LFM2 architecture to achieve high inference speeds without the massive memory overhead typical of parameter-heavy transformers.
While major AI companies Anthropic, OpenAI, Google, Microsoft, Meta and others push parameter counts into the hundreds of billions or trillions to achieve frontier performance, a parallel race focuses entirely on the edge and local deployments.
Liquid AI’s launch of LFM2.5-230M signals a pivotal shift toward architectural efficiency over brute-force scaling. By squeezing 19 trillion tokens of pre-training into a 230-million-parameter footprint, the company demonstrates that edge devices do not need massive computational power or persistent cloud connections to execute complex, multi-step agentic workflows.
The LFM2.5-230M model diverges from standard transformer architectures, relying instead on the LFM2 framework. This architecture functions as a hybrid system, interleaving gated short-range convolutions with grouped-query attention to process information efficiently.
For those tracking the evolution of efficient architectures, Liquid’s approach shares a similar conceptual goal: managing long contexts and sequential data effectively on edge hardware without the quadratic memory costs of pure attention mechanisms. The model supports an expansive 32K context window, allowing it to ingest substantial documents or continuous streams of robotic telemetry.
When analyzing the performance charts provided in the release, the architectural efficiency becomes visually apparent. The model maintains a memory footprint of under 400MB while achieving prefill and decode speeds that outpace comparable models like Gemma 3 1B IT and Granite 4.0-H-350M.
On a Samsung Galaxy S25 Ultra equipped with a Qualcomm Snapdragon Gen4 CPU, the model reaches a decode speed of 213 tokens per second. Even on a highly constrained Raspberry Pi 5, the model maintains a decode rate of 42 tokens per second. Furthermore, internal benchmarking shows the GPU inference stack delivers lower end-to-end latency than competing small models across all concurrency levels.
To understand why a 230-million-parameter model is necessary, one must look at how enterprises currently manage data.
Organizations have traditionally relied on rigid, rule-based Extract, Transform, Load (ETL) scripts to move and process data. However, these legacy systems are notoriously brittle; a simple change in a document’s layout or a schema update can break the entire pipeline.
To solve this, the industry is shifting toward “AI ETL,” where machine learning infers mappings, detects schema drift, and adapts to changes automatically. In a modern lightweight data extraction pipeline, an AI model connects to unstructured sources—like PDFs, emails, or web forms—and structures the data into formats like JSON without requiring hardcoded rules.
For enterprises, using a massive flagship model like Claude Opus 4.6 (which costs $5.00 per million input tokens) to parse routine invoices, format addresses, or route telemetry data is economically unviable.
This is where models like LFM2.5-230M become critical. Designed explicitly as a lightweight extraction engine, it allows companies to automate repetitive formatting and data parsing at a fraction of the compute cost and latency, running directly on local hardware rather than relying on expensive, continuous cloud API calls.
The AI industry in mid-2026 is seeing a renaissance in “small” models, but the definition of “small” varies wildly.
Recently, the open-weight community was stunned by Weibo’s VibeThinker-3B, a 3-billion-parameter model built on a Qwen2-style backbone that achieved a massive 94.3 on the AIME 2026 math benchmark, rivaling 600-billion-parameter behemoths through aggressive data curation and reinforcement learning.
Similarly, Google’s Gemma 4 family — which recently crossed 200 million downloads — pushes frontier AI to the edge, including the E2B (2 billion parameters) designed specifically for mobile and IoT deployments.
By contrast, Liquid AI’s LFM2.5-230M operates in a completely different weight class. At just 230 million parameters, it is roughly one-tenth the size of Google’s smallest Gemma 4 model and VibeThinker-3B.
Because of its microscopic footprint, LFM2.5-230M is not designed to compete on reasoning-heavy workloads like advanced math, coding, or creative writing—a constraint Liquid AI explicitly acknowledges.
However, in its intended domains of data extraction and tool calling, the model punches well above its weight class.
Benchmarks released by Liquid AI show LFM2.5-230M scoring 43.26 on the BFCLv3 tool-use benchmark, dominating IBM’s Granite 4.0-350M (39.58) and completely outpacing larger 1-billion-parameter models like Google’s Gemma 3 1B IT (16.61).
On CaseReportBench for data extraction, it scores 22.51, decimating the Qwen3.5-0.8B (Instruct).
LFM2.5-230M proves that while 3-billion-parameter models like VibeThinker are solving advanced calculus, a 230-million-parameter model is the superior, highly optimized choice for executing structured tool calls and keeping agentic pipelines running efficiently on constrained hardware.
Because it excels at tool calling, LFM2.5-230M functions primarily as a skill-selection layer. Liquid AI demonstrated this capability by deploying the model on a Unitree G1 humanoid robot.
Running entirely on-device via the robot’s onboard NVIDIA Jetson Orin compute module, the model successfully processes complex environmental commands.
As noted in the company’s technical blog, the model takes a free-form instruction like, *”Hold still for 2 seconds, then walk forward at 1 meter per second for 3 meters, hold a forward one-leg kneel for 5 seconds, and walk backward at 0.5 meters per second for 3 meters,”* and automatically translates it into a structured multi-step plan calling on pre-trained low-level skills provided by NVIDIA’s SONIC framework.
The base and post-trained models are available immediately on Hugging Face, with native day-one support across the inference ecosystem for llama.cpp (GGUF), MLX, vLLM, SGLang, and ONNX.
Liquid AI ships LFM2.5-230M under the LFM Open License v1.0. Despite the word “open” in the title, this is not an Open Source Initiative (OSI) compliant license; it operates as a restricted, dual-use commercial framework.
For independent developers, researchers, and early-stage startups, the license functions identically to open-source software.
Users receive a perpetual, worldwide, royalty-free license to reproduce, modify, and distribute the model, provided they retain original copyright notices and prominently state any modifications.
However, the license includes a strict “Commercial Use Limitation”. Any legal entity generating $10 million or more in annual revenue loses the right to use the model commercially under this agreement.
Large enterprises crossing this financial threshold must negotiate a separate, paid commercial agreement with Liquid AI to deploy the model in production.
This strategy protects the company from having its intellectual property absorbed by major technology conglomerates for free, while still seeding the model at the grassroots developer level.
Forget smart tractors. Discover how localized food infrastructure networks like Local Dutch use demand data to bypass broken long-haul logistics
After months of waiting the new Dota 2 Dark Carnival event is finally here and it looks like one of the best in game events ever.
Samsung has confirmed the Galaxy S27 Pro exists via the GSMA database, alongside the S27, S27+ and S27 Ultra. Here’s what “Pro” really means.
A new Escape From Tarkov event has kicked off, giving you the chance to get some rare loot on Woods, but the scavs might have something to say about that.
For bigger engineering teams, the challenge isn’t to copy small-team practices wholesale but to adapt the underlying habits that make them so effective.
OpenAI has made a significant update to its most widely used language model, GPT-5.5 Instant, which is the default in the free version of ChatGPT.
The company announced the upgraded version of GPT-5.5 Instant yesterday on X, calling it “much more fun to talk to” and saying it is “better at understanding the intent behind a question and adapting its response accordingly,” as well as offering improvements in shopping results, local recommendations, and handling “complex constraints.”
However, it has not yet provided any benchmarks or numerical results to quantify these claims.
The company said the updated GPT-5.5 Instant was rolling out first to paid ChatGPT subscribers and then to free users as of today, June 25.
OpenAI also updated its chat-latest API alias, which points to the latest GPT-5.5 Instant model currently used in ChatGPT, while continuing to recommend the separate gpt-5.5 model for production API usage.
That distinction matters, but it should not obscure the main news: this is primarily a ChatGPT-side update to GPT-5.5 Instant, not a new release of the broader GPT-5.5 API model family.
Let’s dig into what’s changed…
GPT-5.5 Instant was first unveiled in early May 2026, just under two months ago, to replace the aging GPT-5.3 Instant engine as the baseline default model for ChatGPT users.
Developed as a fast, high-throughput variant of OpenAI’s core flagship model family, the initial spring release focused heavily on correcting systemic factuality deficits.
Internal benchmarks from that spring deployment reported a 52.5% reduction in hallucinated claims compared to GPT-5.3 Instant on high-stakes medical, legal, and financial prompts, alongside a 37.3% drop in factual error rates on user-flagged historical conversations.
Independent evaluators noted that its predecessor, GPT-5.3 Instant, had struggled in public rankings, placing 44th overall in Arena benchmarks. That gave the May rollout a clear purpose: OpenAI needed a stronger default model for everyday ChatGPT interactions, not just a more capable frontier model for advanced users.
Stylistically, the initial spring model introduced a sharper conversational baseline, demonstrating a 30.2% reduction in word count and a 29.2% drop in line usage over typical advice prompts.
However, the spring deployment also introduced an operational fault line for enterprise software systems: a feature known as “memory sources.” Designed to grant users visibility into the specific past chats, files, and connected Gmail accounts shaping a personalized answer, memory sources introduced a loose, model-reported observability layer.
As reported by VentureBeat, these internal summaries frequently clashed with the deterministic logs of localized vector databases and enterprise Retrieval-Augmented Generation (RAG) pipelines.
The resulting friction created dual, competing context records, making it difficult for administrators to reconcile what the model claimed it referenced against what it actually accessed in production.
The June 24 update does not appear to expand memory sources directly. Instead, it focuses on making GPT-5.5 Instant better at understanding user intent, carrying context across turns, following multi-part instructions, and producing more useful shopping and local recommendations.
For everyday users of ChatGPT, the most noticeable change in GPT-5.5 Instant will be the model’s improved intent recognition.
According to OpenAI’s latest release notes, GPT-5.5 Instant has improved at identifying the underlying goal behind a user’s question, particularly in decision-support scenarios like planning, shopping, asking for advice, researching options and comparing local choices.
Historically, large language models have struggled when given prompts with multiple overlapping constraints — often dropping one or two requirements in favor of a generalized response.
The updated GPT-5.5 Instant handles these complex instructions more reliably. When users push back on an answer, clarify their meaning, or introduce new constraints mid-conversation, the model should adapt dynamically rather than stubbornly repeating its original approach.
This contextual awareness extends heavily into commerce and local recommendations. GPT-5.5 Instant now makes better use of location context to surface nearby options, weaving together product recommendations, business information, and relevant images into a more cohesive output when those elements are useful.
Furthermore, OpenAI notes that the stylistic formatting of these responses is less rigidly templated, trading robotic lists for a more intentionally designed, warmer and restrained conversational tone.
chat-latestFor the developer ecosystem, the June 24 GPT-5.5 Instant update is accessible through OpenAI’s updated chat-latest API alias.
chat-latest is not the same thing as the production gpt-5.5 model slug. OpenAI says chat-latest points to the latest Instant model currently used in ChatGPT, and it recommends the separate gpt-5.5 model for production API usage. Developers can use chat-latest to test the newest ChatGPT-style improvements, while using gpt-5.5 when they need a stable production target.
The current chat-latest model page lists a 400,000-token context window and support for up to 128,000 maximum output tokens. Its knowledge cutoff is Aug. 31, 2025.
On pricing, chat-latest uses the same $5.00 per 1 million input tokens and $30.00 per 1 million output tokens listed on its model page. Cached inputs cost $0.50 per 1 million tokens, a 90% discount that strongly incentivizes developers to optimize prompts by placing static instructions first and dynamic data later.
The model supports text and image input, text output, streaming, function calling and structured outputs. Through the Responses API, the chat-latest page also lists support for web search, file search, image generation, code interpreter and MCP.
The practical takeaway is simple: chat-latest gives developers access to the updated Instant-style behavior, but OpenAI is still steering production API builders toward the separate gpt-5.5 model. The broader GPT-5.5 API model includes a larger feature set and different production profile, but that is not the main focus of this update.
For enterprises, the June 24 GPT-5.5 Instant update lands at the intersection of two related but distinct trends: better default user experience in ChatGPT, and more reliable orchestration behavior in the API.
The consumer-facing changes make ChatGPT more useful for everyday decision-making. Users should see better handling of messy, real-world requests: planning a trip with several constraints, comparing products, finding nearby businesses, or adjusting a recommendation after adding a new requirement.
The enterprise relevance is less about a new technical architecture and more about default behavior. A model that better infers intent, preserves context across turns and follows multi-part constraints can make ChatGPT more reliable for employees using it for research, planning, purchasing decisions, customer-facing drafts and internal analysis.
But enterprises should remain careful about observability. Memory sources can help users understand why ChatGPT personalized an answer, but they do not provide a complete audit trail. Organizations that already rely on RAG pipelines, vector databases, orchestration logs and internal agent traces should define which record acts as the source of truth when a model’s visible memory sources do not fully match the system’s own logs.
The release of GPT-5.5 Instant and the updated chat-latest alias signals a maturation in how generative models are deployed.
OpenAI is moving away from models that require heavy hand-holding and toward systems that can better infer the user’s goal, preserve constraints and adapt across multiple turns.
Whether it is a consumer planning a complex multi-city vacation in ChatGPT, or a developer orchestrating a codebase-navigating agent through the API, GPT-5.5 represents a faster, smarter and more capable baseline for the future of AI workflows.
The most important takeaway for developers is also the simplest: GPT-5.5 Instant, chat-latest and gpt-5.5 are related, but they are not the same product surface. GPT-5.5 Instant is the ChatGPT model users experience directly. chat-latest is a moving alias for testing the latest Instant behavior through the API. gpt-5.5 is the production model OpenAI recommends for developers building stable applications.
For decades, the enterprise technology industry operated on a simple principle: software companies built products, and services firms helped enterprises.
Samsung’s health features are growing apace, and will offer insights through wearables and other devices, the company’s health specialist said.