Broad AI assistants get the attention, but the launches that tend to stick are the ones that do exactly one job for one group of people. Here are seven from the last few weeks, each narrowly aimed and each live today.
Compare small models on your own GPU
MicroLLM Lab runs quantised small language models entirely in your browser with WebGPU. You load a model, chat with it and benchmark it against the rest of the catalog, with no server, no account and nothing uploaded. It is the fastest way to answer a question that published benchmarks cannot: how does this tiny model actually feel on my hardware. The catalog covers models in the same range as PetitGPT and SmolLM2.
Chart with your own agent
OpenChart is an open-source trading workspace built as a TradingView alternative that brings your own agents into the charting workflow. Instead of a closed platform with its own scripting language, you get an OSS workspace where the agents you already run sit alongside the charts. Analysis and automation end up in the same place, which is the part commercial platforms tend to keep separate and charge for.
Analytics that can refuse a query
Strata pairs a governed semantic layer with dashboards designed for self-service analytics that stays safe when AI and agents query it. It routes queries across federated Open OLAP engines with partition-aware routing to hot tiers, and the semantic layer is built to refuse a request rather than answer it badly. That refusal behaviour is the whole thesis: an analytics layer that always answers is one bad prompt away from confidently wrong numbers. A live demo runs in the browser or through Docker, with no credit card.
Edit Word documents without breaking them
Vespper ships DOCX MCP, a model fine-tuned specifically for editing Word documents and exposed as an MCP server. General agents struggle with tracked changes and formatting, and Vespper targets the contracts and regulatory documents where a Word file is the actual deliverable. The team reports three times faster and two times cheaper edits on its internal benchmark, which is a vendor number rather than an independent result, but the specialisation itself is the interesting part.
Edit photos without a subscription
Rembrandt is a free photo editor with RAW support, masks and on-device AI for macOS, Windows and Linux. There is no account, no subscription and no tracking, and it can be self-hosted. It suits photographers who want a Lightroom-class workflow without handing a library to a cloud service. On-device AI is the detail that makes it viable: edits do not need to leave the machine to be useful.
Voice in both directions
Nari Labs offers Qwen3-TTS and Qwen3-ASR as text-to-speech and speech-to-text APIs aimed at accuracy, latency and cost. The team reports leading Coval's September 2026 benchmarks on STT latency and TTS word error rate, sitting on the quality-latency Pareto frontier for both directions. One practical note: the free public beta ended in September 2026, so ongoing use now goes through GA model IDs rather than the beta endpoints.
Build LEGO models by describing them
LDraw Nova is agent tooling for generative LEGO model building, built with Astra and Opus 5.5 and powered by Jev. Rather than placing bricks by hand, you describe what you want and the agent assembles an LDraw model, which is the open file format the LEGO CAD community already uses for parts, builds and instructions. Outputting the format builders already work in, instead of a proprietary one, is why this has a chance of being used rather than just demoed.
Why narrow keeps winning
There is a reason this pattern keeps repeating. A general assistant competes on breadth, which is the one dimension where the largest labs have an insurmountable lead. A tool that edits Word files well, or renders LEGO builds in the format builders already use, competes on depth inside a market small enough that nobody with a bigger model bothers to enter. The constraint is not a limitation, it is the moat.
The pattern
Each of these seven picked a narrow job and built the whole product around it: a specific model size, a specific file format, a specific refusal behaviour, a specific domain. That is a better bet for a small team than competing with a general assistant, and it is where most of the genuinely new launches are coming from this month.
