Fetching from the wire…
Models2026-09-13 · source-backed
yifanzhang-pro/recurrent-looped-tranformer was created September 12 at 19:02Z and had 341 stars by the following midday. It pairs a causal encoder with a recurrent decoder carrying its final hidden state and layerwise sliding-window attention cache across every prompt and response token, 48 encoder layers and 48 decoder layers with attention and FFN weights shared across stages. The project page says outright it's a specification rather than an empirical report, and that "realized reasoning gains, hardware efficiency, and RL scaling remain to be established." No arXiv version. The stars are for the architecture diagram.
Each link below shares sources, entities, or timing with this story.
Qiuner/birdview was created September 12 at 06:59Z, MIT, JavaScript, pitched as "stop letting AI code blind" by mapping a codebase's architecture before every change. The 40:1 star-to-fork ratio on a day-old repo points at attention rather than adoption, and there's no seconda...
The model card went up at 02:17 UTC on September 10, MIT-licensed, multimodal MoE with a 1M-token context. The Causal Encoder-Decoder architecture projects the decoder's global KV cache from final encoder hidden states rather than per-layer, and SWA Bounded Replay cuts persist...
Build b10758 (September 2) fuses matmuls landing on Hexagon HMX, fuses MUL_MAT_ID into MUL_MAT_ID_NX, and adds VA defragmentation so large-dim runs abort less on fragmented address space (release). Build b10757 handles batch sizes above 4 for IQ3_S mat-vec when NUM_COLS > 4, r...
v0.8.35 released September 11, Apache-2.0 TypeScript, 4,485 forks, 133 open issues, created March 10, so 49K stars in six months (GitHub). The premise is that an agent already writes HTML and GSAP competently, so making HTML the video authoring format removes the need for a vi...
H Company released it September 3, a multimodal-native encoder at 260M and 800M running text tokens and raw image patches through one bidirectional Transformer with no separate vision tower or text decoder, using a from-scratch 131k BPE vocabulary, 16,384-token context, 32x32...
zendev-sh/goai (Go, 194 stars, 21+ providers, stdlib only), juspay/neurolink (TypeScript, 128 stars, 24+ providers), sno-ai/llmix (Python, 132 stars, adding cache, retries, circuit breakers and key rotation) and r-uby-dev/llm (Ruby, 139 stars, A2A and MCP client), all pushed b...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.