Daily Briefing
5 August 2026 · 5 stories
🔥 Top story
01An AI agent that fits in your pocket: Liquid AI releases a 2.6 billion-parameter brain, free to download
Picture an assistant that does not talk to a remote server: it thinks inside your phone, using less than 2.5 GB of RAM. That is what Liquid AI released with the LFM2.5-2.6B model, trained specifically to drive agents — so it can pick the right tool, break a task into steps, call other software, and assemble the answer. The model runs at 220 words per second on a high-end Mac and 30 words per second on a regular smartphone. On the published benchmarks, it outperforms models three to four times larger on instruction-following and tool-use. The free weights are available on Hugging Face right now, for curious tinkerers and for small businesses that do not want to depend on a cloud API.
02Ant Group opens a Chinese ultra-light model (124 billion parameters, only 5 active) under an MIT license
InclusionAI, the open-source branch of Chinese giant Ant Group, released the full weights of Ling-3.0-flash on Tuesday. The paradox: 124 billion parameters in total, but only 5 billion active on any single question — an ensemble of experts that wakes up on demand. Hybrid architecture, 262,000-token context window, MIT licence with no restrictions. Stated target: agentic workloads (using tools, following a multi-step plan, chaining software calls). It plugs natively into the harnesses developers use to build agents: Claude Code, Hermes Agent, OpenClaw. In plain terms: a new free playground for anyone who wants to run agents over long sequences, without a proprietary lock-in.
03Your smartphone camera becomes an autonomous visual agent: NewEyes scans, reasons, and acts
Point your phone at a scene, and the NewEyes app does not just describe what it sees: it chains actions. Scan your wardrobe to tell you what is missing, read a parking sign and remind you to move the car, turn a hand-drawn sketch into a website before you leave the room. That is the bet from Collov Labs, which launches its first consumer product since its 23 million dollar Series A in early 2026. The processing runs partly on the phone's dedicated neural chip (Qualcomm Hexagon), so nothing has to be sent to the cloud. It already works with Meta Ray-Ban smart glasses in beta. For anyone who finds text agents a bit dull: AI that acts in the physical world starts through the eye.
04Nimble promises web search agents that cut token use in half
When an AI agent spends its life searching the web to answer its user, the bill adds up fast. Nimble offers a search agent that learns the company's business domain (sales, marketing, finance, support) and adapts its search strategies accordingly. Headline result: less than half the tokens consumed per query, with answer quality up by 21 points. First named customer: Rox, an AI-native CRM, which reports a twenty-fold reduction in agent costs after adoption. Connectable as a tool via a standard API or an MCP connector. For small and medium businesses scaling up agents, it is a concrete path to a smaller monthly bill.
05An AI neocloud pockets 300 million to become the GPU lender to small startups
Volta Infra Holdings stepped out of stealth on Tuesday with a 300 million dollar equity raise, a 5 billion dollar structured financing envelope, and a 10 billion dollar six-year partnership to build an AI compute centre in Norway. On the cap table: Andreessen Horowitz and Altimeter as co-leads, plus Nvidia and Michael Dell's family office. The target: 'little tech' — those AI startups that need a lot of compute but do not have the balance sheet to sign a multi-year deal with Microsoft or Amazon. Volta plays the middleman: it buys Nvidia chips, finances the data centre, then resells capacity by the hour or by the month. For European labs struggling to find GPU slots, it is one more sign that infrastructure is diversifying — beyond the three cloud giants.
📡 To watch
Voluntary cyber-test framework: the White House meets OpenAI, Anthropic, Google and Meta
Meeting on 4 August to finalise a protocol for early government access to frontier models. The official text has not been published yet — watch for it this week.
Ling-3.0-flash numbers: the stable variant lands this week
Published scores come from a 'release candidate' with thinking mode enabled. InclusionAI should publish a table linking those numbers to the final model served on OpenRouter.
NVIDIA reports quarterly earnings on 26 August
The Volta/Nvidia dynamic will be tested: the company is betting big on GPU verticalisation towards neoclouds. Watch the data-centre guidance.
Trump-Xi summit in September: possible pivot on Chinese open-weight models
Cited by several analysts (Samm Sacks, New America) as a possible inflection point on US restrictions against Chinese labs.
📊 Trend
Two undercurrents cross this week. On one side, the 'agent-as-black-box' chain sets a new methodological standard: Liquid AI runs its agentic reinforcement learning by treating agent harnesses (Claude Code, Hermes Agent, OpenClaw) as black boxes whose trajectories are captured without modification. InclusionAI follows the exact same logic on the integration side. On the other side, the building permit: Volta raises 300 million to buy Nvidia GPUs on behalf of young labs, while the White House regulates access to frontier models. For agent builders, the lesson is clear: reasoning compresses (models fit in a pocket), harnesses become the reference target, and physical infrastructure becomes a governance topic in its own right. Keep all three windows in mind for the next six months.