On July 31, 2026, new locally runnable LLMs dropped one after another, while frustration mounted in the community over models that tout themselves as "open" yet never release their weights. From subtle moves in small and mid-size models, to a new security risk around AI agents, and even a corporate AI pullback, we've rounded up the stories that caught our eye into six topics.

256k Context via a 30B Lookup Table Kept in RAM — Meituan's LongCat-Flash-Lite-Sparse

China's Meituan has released [LongCat-Flash-Lite-Sparse](https://huggingface.co/meituan-longcat/LongCat-Flash-Lite-Sparse). It's a Mixture of Experts (MoE) model with around 3B active parameters. The attention-grabbing part is the design that offloads a 30B-scale n-gram lookup table to RAM, reportedly allowing a 24GB GPU to process 256k context at high speed.

Early analyses include remarks that it "brings to mind the PLE (parameter allocation) technique from Gemma 4," and it's regarded as a clever way to stretch limited VRAM. On the flip side, more cautious assessments say it "isn't a replacement for the Qwen 3.6 27b you already run," suggesting it's best seen not as an all-purpose flagship but as a supplementary model specialized for long-context work. The stream of open-weight releases from Chinese players shows no sign of slowing this week, and eyes are also on what MiniMax does in the week ahead.

Ling 3.0 flash Builds a 3D City from a Single Prompt

Also generating buzz on LocalLLaMA was a spatial reasoning test of Ling 3.0 flash. Working through MCP (Model Context Protocol) to drive Blender, the model wrote its own Python and constructed a city complete with elevated highways, supertall buildings, and material settings. It even set up a camera path and rendered an aerial flythrough video.

Unlike one-shot image generation, the model was praised for completing a task that demands long-horizon tool use and spatial planning. For now it's free to try via OpenRouter, and an open-source release with vLLM support has been teased for the near future — a development worth watching.

"Open" but No Weights to Be Found — Proposal to Introduce a [no weights] Tag Draws Attention

Even as open-weight models arrive one after another, what's fueling community frustration is the pattern of models that promise a release and then never deliver the weights. What was raised on LocalLLaMA is a rule proposal to explicitly tag such models as [no weights].

The poster pointed out recent cases where the weight release slipped to "a few weeks away," and cases where nothing was ever published at all (calling out Meta by name). Since the forum's core topic is "LLMs that run locally," the motivation is to make these easier to distinguish from "maybe local someday" models. Whether the idea catches on depends on how it's put into practice, but it's a symbolic debate that mirrors the gap between expectation and reality in the open-weight world.

"Stealthy Audio Prompt Injection" Targets Multimodal LLM Agents

On the security front, what stood out was the arXiv paper [Stealthy Concurrent Audio Prompt Injections Against Multimodal LLM Agents](https://arxiv.org/abs/2607.28165).

It appears to examine methods for launching audio-based prompt injection against multimodal LLM agents that handle audio, in a hard-to-detect (stealthy) fashion. As AI agents increasingly process external input through microphones and calling features, the audio attack surface is becoming a very real concern. For the details, the paper itself is the place to go, but if voice AI continues to spread, this is a direction that can't be ignored.

Starbucks Axes Its Nationally Deployed AI Inventory Tool After 9 Months

A case study in the harsh realities of enterprise adoption. Starbucks bet on NomadGo, an AI tool that automatically counts inventory, and deployed it at national scale — but reportedly pulled the plug just nine months after rollout.

Despite the splashy adoption announcements, it seems the tool may have fallen short in on-the-ground accuracy and day-to-day operations. As an example that forces a rethink of the criteria for moving an AI tool from PoC (proof of concept) to full-scale deployment, it's an instructive episode for anyone weighing adoption.

Running Qwen 35B on an RTX 4070 Averages 42W — Measuring Power Draw During Inference

Lastly, some real-hardware numbers. On Qiita, power consumption profiles for running Qwen 35B on an RTX 4070 were measured across four patterns.

The takeaway: average GPU power draw during text generation (decode) was 42W. The card's TGP (Total Graphics Power) is 200W and gaming averages 205W, while the peak during prompt evaluation (prompt-eval) briefly spiked to 175W. The numbers demonstrate that inference keeps power draw far below gaming-load levels — encouraging data for local operators who care about electricity bills and heat.

---