Fetching from the wire…
01
02
03
04
05
06
07
08
Models2026-08-11 · source-backed
Cactus Compute released it August 10 under Apache 2.0: 28 MB peak session RAM, 70 MFLOPs per token via CQ2-bit compression, 800+ tok/s prefill and 500+ decode on a Raspberry Pi 5. Targets span Cortex-M microcontrollers through x86 and WebAssembly. The honest limits are published too: on the 961-row Mobile Actions benchmark it hits 98.3% function-name accuracy but 71.3% single-call and only 48.4% two-call. Single-shot tool dispatch on hardware where nothing else fits, not multi-step reasoning.
Each link below shares sources, entities, or timing with this story.
Can a model that fits on a Raspberry Pi do reliable tool calling? Two independent labs just answered yes. PrismML emerged from stealth March 31 with Bonsai, the first commercially viable 1-bit LLMs built on Caltech research. The 8B model fits in 1.15GB (vs 16GB for FP16), runs...
Agentic Resource Discovery went up August 24 at agenticresourcediscovery.org under Apache 2.0, contributed to but not authored by AWS (AWS ML Blog). It lets agent, tool and skill registries federate across clouds, on-prem and SaaS without bilateral connectors. AWS Agent Regist...
The August 14 report covers January through August 2026: model repos grew from 2.43M to 2.96M, datasets from 711K to 1M, and 85.6% of models have under 200 lifetime downloads (Hugging Face). Chinese labs shipped monthly parameter ceilings of 754B to 2.78T against sub-130B for...
Released under Apache-2.0 on August 13, NAC targets the context-rot failure mode where investigation, debugging and false starts all pile into a single transcript until the model degrades (Arcee AI). A central orchestrator plans but cannot execute, fresh worker processes run t...
~248 stars today, 3,921 total, MIT (GitHub). The size is the entire argument, at 14MB it fits a class of device where even quantized SLMs are impractical. Last shipped August 11, no formal releases, so treat every capability claim as unbenchmarked until someone independent run...
Carloscodix/qapla hit the HN front page at 41 points one day after creation, 43 stars, Apache 2.0. The explicit claim is training, not inference: the complete loop with backpropagation by hand on an ESP32-S3. A curiosity, not a product. But it's a clean floor estimate for trai...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.