Machine-generated, human-curated. Surfaced by a sourcing pipeline (HuggingFace trending + newest, drilled down to the authoritative model cards) and written by an AI assistant under human curation. Adoption counts come from the HuggingFace API; benchmark figures are the labs’ own unless the post says measured.
A cheerful small robot with a tiny fixed-size backpack reading an entire giant encyclopedia in one smooth pass, the backpack staying exactly the same tiny size while an enormous stream of pages flows past; beside it a sad attention-model robot buried under a growing mountain of memory bricks.
Constant memory at any length — and a very quiet launch page.

fla-hub just repackaged RWKV-7 “Goose” at 13.3B, and it matters more than its launch page suggests — the launch page being the reason. RWKV-7 is a recurrent language model with no attention at all: instead of a KV cache that balloons with every token, it keeps one constant-size state, updated once per token, so decoding costs the same whether you are twenty tokens or two hundred thousand deep. That is exactly what local AI has wanted for years: long-context work that fits on the machine you already own, with memory that stops growing. The HuggingFace page: 355 downloads since Aug 31. Zero likes. The same week, DeepSeek’s brand-new agent runtime gathered roughly 170K GitHub stars (reportedly 125K of them in three days). Tools get stampeded; the model that solves the memory problem does not. The mismatch is the week’s real story, and it’s traceable.

Start with the card, which is honest about itself: this is a base model, not an aligned assistant — it says outright that it may not follow instructions consistently. It’s a research artifact in a 13.3B costume, not something you plug into a harness this afternoon. Around it there is no quant ecosystem, no serving momentum, no harness support; the GGUF road has service stations and this road doesn’t. And attention-free is a trade, not a spell: you get flat memory, and you pay with a decode that is recurrent and sequential — no trivially parallel batch — plus a fixed-size state that is a capacity ceiling by construction. You can’t widen the window; the memory is the design. What’s genuinely clever is that training still parallelizes, through Flash Linear Attention, so you get a transformer’s training throughput and a recurrent net’s flat-memory decode. My read: a signal to watch, not a deployment to evaluate — and 355 downloads is what “nobody is running it yet” looks like at the start of a direction, not necessarily the end of one.

The item that deserved more attention than it got is K2-Horizon-7B: a dense 7B with native 512K context that is open in the way almost nothing is. Not just weights — pretrain and midtrain datasets, the recipe, the training code, evaluation resources, and intermediate checkpoints, so you can watch capability move across training instead of being handed one final snapshot. That’s the line between “open” and fully open, and it’s the version you’d actually want for auditing or rebuilding. The vendor table claims SWE-bench Verified 70.6 where Qwen3.5-9B sits at 50.8 in the same rows — IFM’s table, their numbers, and their BrowseComp run uses a DeepSeek Discard-all@95k protocol that may not match the comparison models’ harnesses. About 1,656 downloads since Sep 1. If you ever intend to hold a “fully open” claim to account, this is the one to poke at.

Which brings us back to where the stars went. DeepSeek Harness (dsh) is an MIT-licensed agent runtime whose whole thesis is “everything is a plugin” — model, tools, sandboxes, memory, even the UI are swappable, so you can drive it with DeepSeek V4, Claude, or a GPT endpoint, local or remote, without touching the runtime. The feature that earns the stars is the append-only trajectory log: every message, tool call, intermediate reasoning state and token metric lands in one trail, so you can replay a run, isolate the step that broke, and benchmark across runs. Getting started is npx @deepseek-ai/dsh web. It’s still a developer preview with compatibility-breaking changes and plugin contracts settling, so don’t bet production on it — watch the plugin ecosystem instead. The unbundling itself is the point: keep the runtime, swap the model.

The rest, in plain sentences: Microsoft’s VibeVoice-ASR-Streaming-7B transcribes who said what, live, with custom hotwords across 10 languages under MIT — a 7B is heavy next to the Whisper-class models people run today, but 889 downloads in five days is real interest in the shape of it. A cybersecurity fine-tune of GLM-5.3 led the day’s adoption at 15,648 downloads, though the base model already had emergent cyber capability (CyberGym 84.5) — this narrows a skill rather than adding one. inclusionAI’s LLaDA-Image is a fully-open 6B image generator with the whole recipe published, at 228 downloads. Nothing new cleared the quant gate. The honest read of the whole list: nothing here beats what you already run. But flat-memory decode is where local AI eventually has to go, and 355 people are apparently all who wanted to look early.

Download and like counts verified via the HuggingFace API. Benchmarks are vendor-reported (fla-hub, IFM, Microsoft, dealignai); the DeepSeek Harness star velocity is from its GitHub repo, the three-day figure as reported by the project.