Releases·Tool Release
Cognitive Sharding: Computer-Use Agents Run Locally on 16 GB via Specialist Models
A developer published Cognitive Sharding, an architecture that splits computer-use agents across three specialist local models coordinated by a code-owned control plane, fitting the full stack in 16 GB of memory. It prioritizes accuracy over latency by verifying the environment after each step to contain errors in long workflows.
6 engagement2 sourcesFri, Sep 25, 2026, 01:47 AM
AI Overview
- Off-Grid AI's Cognitive Sharding splits computer-use agents across three local specialists: Bonsai 2 27B for planning, Kev 4B for action selection, UI-Mate 9B for visual grounding
- A code-owned control plane schedules model residency within 16 GB, maintains task state, restricts per-model actions, and verifies the environment after every step
- The architecture trades latency from model swaps and verification for error containment across workflows of hundreds of actions
- Off-Grid AI argues capable local agents depend on how cognitive functions are composed, not solely on single-model capability, and seeks comparisons for 300+ action workflows
Entities
Related
Developments·Paper
OS-Helix: end-to-end trained 27B/35B agent unifying GUI actions and code execution, hitting 66.44% on OSWorld-Verified
Researchers introduce OS-Helix, an end-to-end trained multimodal computer-using agent that unifies GUI actions and code execution in one action space. Naively adding a code tool degrades GUI agents, so GUI/code routing is internalized via SFT and online RL on hybrid desktop data. OS-Helix-27B-RL reaches 66.44% success on OSWorld-Verified, a 42.9% relative gain over its baseline.
Researchers introduce OS-Helix, an end-to-end trained multimodal computer-using agent that unifies GUI actions and code execution in one action space. Naively adding a code tool degrades GUI agents, so GUI/code routing is internalized via SFT and online RL on hybrid desktop data. OS-Helix-27B-RL reaches 66.44% success on OSWorld-Verified, a 42.9% relative gain over its baseline.
0 engagement1 sourcearxivThu, Oct 15, 2026, 05:14 PM
Products·building
SIA: local agent runtime that gets more out of small LLMs (3B–8B)
SIA is an agent runtime for local LLMs, built on the premise that a better system around a small model can substitute for a larger one. It targets locally hosted 3B–8B models and is being developed against Qwen2.5 Coder Tools 7B. It serves hobbyists who want to experiment with AI agents on consumer hardware that can't run large models.
SIA is an agent runtime for local LLMs, built on the premise that a better system around a small model can substitute for a larger one. It targets locally hosted 3B–8B models and is being developed against Qwen2.5 Coder Tools 7B. It serves hobbyists who want to experiment with AI agents on consumer hardware that can't run large models.
13 engagement1 sourceredditTue, Oct 6, 2026, 08:25 PM
Social·Community
Users experiment with Claude to orchestrate small local models for coding tasks
Reddit users are exploring hybrid setups where a cloud-based frontier model (like Claude or Fable) orchestrates smaller local models (e.g., Qwen3.6-27B, 14B, 1.5B–3B) to handle sub-tasks, aiming to reduce API costs and usage limits. One user successfully used Claude to design an experiment where a 14B model wrote specs and 1.5B–3B models implemented code pieces, all running locally, solving a well-structured programming problem. Others are seeking tools and harnesses to efficiently delegate narrow implementation and codebase exploration to local models.
Reddit users are exploring hybrid setups where a cloud-based frontier model (like Claude or Fable) orchestrates smaller local models (e.g., Qwen3.6-27B, 14B, 1.5B–3B) to handle sub-tasks, aiming to reduce API costs and usage limits. One user successfully used Claude to design an experiment where a 14B model wrote specs and 1.5B–3B models implemented code pieces, all running locally, solving a well-structured programming problem. Others are seeking tools and harnesses to efficiently delegate narrow implementation and codebase exploration to local models.
Updated Wed, Jul 22, 2026, 10:08 AM — Users report using Claude to orchestrate local models (14B, 1.5B–3B) for coding, with one successful experiment; others
30043 engagement3 sourcesredditSun, Jul 19, 2026, 09:09 PM
Developments·Paper
Neurosymbolic agent combines VLM exploration with PDDL planning for household tasks
A new arxiv paper proposes a neurosymbolic agent that factors long-horizon household tasks into task-directed visual exploration and constrained symbolic planning. The approach uses a vision-language model to ground predicates and instances, then a PDDL transition model for planning, aiming to improve executability over pure language-model plans.
A new arxiv paper proposes a neurosymbolic agent that factors long-horizon household tasks into task-directed visual exploration and constrained symbolic planning. The approach uses a vision-language model to ground predicates and instances, then a PDDL transition model for planning, aiming to improve executability over pure language-model plans.
0 engagement1 sourcearxivMon, Aug 17, 2026, 04:50 PM
Developments·Paper
Qwen-UI-Agent technical report proposes foundation GUI agent for real-world devices
The Qwen team released a technical report introducing Qwen-UI-Agent, a foundation GUI agent designed to operate across mobile, computer-use, web, and DeepSearch environments. It aims to combine GUI interaction with CLI execution, handle long-horizon tasks, and improve autonomously, advancing toward real-world use.
The Qwen team released a technical report introducing Qwen-UI-Agent, a foundation GUI agent designed to operate across mobile, computer-use, web, and DeepSearch environments. It aims to combine GUI interaction with CLI execution, handle long-horizon tasks, and improve autonomously, advancing toward real-world use.
0 engagement1 sourcearxivThu, Jul 30, 2026, 01:58 PM