Mistro Large 4 uses only part of its model for each token. Its 4-bit weights are still estimated at about 525 GB. Doing less work for a token does not mean the whole model takes less memory to hold.
That distinction helps make sense of several numbers in this story. A benchmark score, a memory figure, a search result, and a menu count each answer a different question. Reading one as an answer to another can make a system look more practical than the evidence shows.
Mistro Runs Part of Its Model, but Must Hold All of It
Mistro Large 4 has 1.05 trillion parameters and uses a mixture-of-experts design. Its router can choose any expert, so all the experts have to stay in memory even though only part of the model runs for each token. The amount of work per token and the amount of memory needed to hold the model are separate limits.
The preview scored 81.7% on CyberGym E2EA, a benchmark with 131 tasks. MIMO V2.6 Pro scored 78.6%, and GPT6 Luna Max scored 77.9%. That is a lead on this benchmark, with context that matters: an agent gets zero for a refusal.
Claude Opus 5.5, GPT6 Astra, and four other models refuse at least 98% of the tasks. Refusal behavior is therefore part of what those CyberGym scores compare. The score tells us something useful about the preview on these tasks, but it does not answer whether the model fits in memory.
Cost is another separate measure. Artificial Analysis puts Mistro Large 4 at $1.13 per intelligence index task, versus 25 cents for GLM 5.3 flash. The preview also wrote 200 million output tokens, compared with a median of 81 million. Mistro says Large 4’s weights will be released by the end of October; the preview’s CyberGym lead still needs to hold when the final model ships.
Uber Can Find a Tool Without Loading the Whole Catalog
Uber reports more than 5,000 MCP tools across more than 800 servers. Its Omni MCP approach has an agent find a server, list that server’s tools, fetch the schema for one tool, and then call it. The agent does not need the whole catalog in its context at once, and the gateway can return only the fields the agent asks for.
Those discovery steps also take work. Does keeping the catalog out of context save enough to justify them? Uber has published no measured token saving or latency result, so its description does not answer that question.
A separate Anthropic system cut tool-related context from 150,000 tokens to 2,000. That is a striking comparison, but it does not measure Uber’s approach. Uber says permissions are enforced when a tool runs; what an agent may see while discovering tools remains unclear.
Phone Video Search Has a Memory Figure and a Retrieval Result
Google says Embedding Gemma 2 uses about 567 megabytes of RAM on a Pixel 11 Pro with quantization. Its image, sound, and text parts put inputs into a shared space, allowing a text or spoken query to match a clip stored on the device. The memory figure makes the phone use case interesting, but fitting in RAM says nothing by itself about which clip a search returns.
For general video search, Google reports a 50.67 hit at one: the right clip ranks first about half the time. Google also says cutting vectors to 256 numbers keeps about 95% of search quality across image, video, and speech. The smaller vectors and the general video search result describe different parts of the proposed use.
The spoken-memo example has a narrower evidence base. The speaker found no source test for searching video with a spoken memo and had not tested it. The reported general video result should stay attached to general video search.
Photocraft’s Menu Count Does Not Show a Ready Editor
Photocraft’s README reports that its PSD resave check passes 307 of 309 files. That check compares what Photocraft renders before and after saving. It does not compare Photocraft’s rendering with Photoshop, so the high pass count answers a more limited question than it may first appear to.
A separate comparison with the flattened image stored in a PSD has an enforced floor of 229 of 309. That is a test floor, not an exact current pass count. The two checks should not be read as interchangeable measures of how accurately Photocraft handles PSDs.
The project reports 626 of 626 Photoshop menu items present. Its own scorecard gives a different view of using them: 60 of 135 settings do nothing, and only three of 47 performance scenarios meet their targets. The menu count says the commands exist. It does not tell us an editing workflow is ready.