A Field CTO PerspectiveGary Berger

Research

AI Research Index

Open-source AI projects I’m tracking, grouped by area: local inference on owned hardware, multi-agent orchestration, coding agents, memory and reasoning, agent tooling, and agent sandboxes.

49 projects · stars as of Nov 4, 2026

Showing all 49 projects

Local inference and edge serving 5

ProjectWhat it isLicenseStarsLinks
Tenstorrent stackTenstorrentWe own a QuietBox 2The open software stack for our Tenstorrent QuietBox 2. tt-metal (TT-NN ops + TT-Metalium kernels) is the core, tt-forge is the MLIR compiler for PyTorch/JAX, tt-inference-server serves models with vLLM, and tt-studio is a web GUI to deploy and chat with models. Latest tt-metal: v0.81.1 (Oct 23, 2026). Latest tt-inference-server: v0.26.0 (Oct 30, 2026) , adding QuietBox 2 support for K2-Horizon-7B, Muse-Glimmer-30B, Kolibri-1-BF16, gemma-4-26B-A4B-it and granite-4.2-30b.Apache-2.01.7k tt-metal, Oct 31, 2026tt-metalInference servertt-forgett-studioDocs
QB2 agentic coding guideTenstorrentNew Oct 31Your QuietBox 2Tenstorrent's Oct 27, 2026 step-by-step: serve Qwen3.8-27B on a TT-QuietBox 2 with tt serve , then connect OpenCode from a Mac with tt launch . One box batches up to 8 agents, and code never leaves your network.Guide (Apache-2.0 tooling) Caution DFlash2 drafter is CC BY-NC-ND (non-commercial)n/a Blog postGuideModelOpenCode docsQB2 docs
colibriJustVuggNew Nov 2, 2026Pure-C, zero-dependency engine that runs frontier MoE models on your own hardware by streaming experts from disk.Apache-2.046.2kGitHub
FreeTokenFlashML-orgNew Oct 31Runs 290B+ frontier MoE models (DeepSeek-V4, Qwen3.6, GLM-5.2) on consumer RTX GPUs; OpenAI/Anthropic-compatible API so Claude Code and Codex can run locally.Apache-2.014.9kGitHubSitePaper
Project Mayamw00New Oct 31, 2026Runs 321B–748B MoE models (GLM-5.3-Flash, DeepSeek-V4.1-Flash, MiniMax-M3.1-Flash) on your own 1–16 GPUs by tiering experts across VRAM, RAM and NVMe; OpenAI/Anthropic-compatible API.MIT Caution Optional DFlash2 drafter is non-commercial (CC BY-NC-ND)2.3kGitHubModels

LLM training 1

ProjectWhat it isLicenseStarsLinks
nanochatAndrej KarpathyNew Nov 1Minimal, hackable end-to-end LLM harness for one GPU node: tokenizer, pretraining, fine-tuning, RL, eval and chat. Trains a GPT-2-grade model in about 2 hours on 8×H100 for ~$48.MIT59.4kGitHubGuideDeepWiki

Open-weight models 3

ProjectWhat it isLicenseStarsLinks
DeepSeekDeepSeekNew Oct 31Frontier MoE models with open weights. Latest Flash is about 763B params with 1M context and native vision; V4-Pro-0813 is the flagship.MITDeepSeek-V4.1-Flash-1017 Oct 17, 2026Hugging FaceLatest modelChangelog
Liquid AI LFMLiquid AINew Oct 31Hybrid, on-device foundation models. LFM3 is six ~2B models for text, vision, audio, omni and decisions on one backbone. Also LFM2.5-24B-A2B (Oct 13, 2026): a 24B MoE with 2B active for on-device agents.LFM Open License v1.0 Caution Commercial use free only under $10M annual revenueLFM3 Oct 27, 2026Hugging FaceLFM3Docs
GLMZ.ai / ZhipuNew Oct 31Top open-weight coding models. GLM-5.3 is a 744B MoE with 40B active and 1M context. GLM-5.3-Flash (Aug 26, 2026) is a 320B MoE with 18B active plus vision, under MIT.GLM-5.3 License Caution Custom; Flash is MIT. Model-as-a-service providers over $10B revenue need Z.ai reviewGLM-5.3 Weights Aug 28, 2026Hugging FaceGLM-5.3GLM-5.3-FlashRelease notes

Multi-agent orchestration 7

ProjectWhat it isLicenseStarsLinks
DeepSeek Super-agentdsh-superagentNew Nov 1One always-on agent across chat, GitHub and email. A filter screens events, a small, fast Router talks to people and hands out tasks, and isolated Workers each own one task, with the strongest model on the hard ones. Long jobs run async; a Watchdog catches stuck workers. In DeepSeek's simulation, p95 tasks finished in 15.7 min vs about 29 min for one session, at 2 to 4x the tokens.MIT Research prototype for DeepSeek Harness, not a product7.0k Nov 1, 2026GitHubArticleHarness
AgentBatonNVIDIA (paper)New Nov 1Learns when to hand a long agent task from a capable model to a cheaper one, once. Cuts API cost 65.9% on average across 11 model pairs while largely keeping accuracy.CC BY 4.0 (paper) Caution Paper only, no code yetn/a PaperPaperPDFSummary
CAMELMulti-agent framework researching the scaling laws of agents.Apache-2.018.0kGitHubDocs
MetaGPTMulti-agent "AI software company" turning natural language into software.MIT71.1kGitHubSite
Atomic AgentsLightweight, modular building blocks for agents.MIT6.3kGitHubDocs
GPTSwarmSelf-improving agent graphs via RL and prompt optimization.MIT1.1kGitHubSite
BurrApacheBuild, trace, and persist decision-making apps and agents on your own infra.Apache-2.02.6kGitHubSite

Coding agents 3

ProjectWhat it isLicenseStarsLinks
SWE-agentTakes a GitHub issue and tries to fix it automatically (NeurIPS 2024).MIT20.7kGitHubSite
OpenHandsAI-driven software development platform.MIT92.3kGitHubSite
Agent ZeroGeneral-purpose personal agent framework.Custom / unclear Caution19.6kGitHubSite

Memory and reasoning 7

ProjectWhat it isLicenseStarsLinks
auto-ontologyNVIDIANew Nov 4, 2026Python framework for semantic, graph-based reasoning layers over data: deterministic query generation and better context retrieval for LLM agents.Apache-2.03.8kGitHub
claude-memthedotmackNew Oct 31, 2026Persistent cross-session memory for coding agents: captures tool use, compresses it, and injects relevant context into later sessions (Claude Code, Codex, OpenCode, Gemini, Copilot).Apache-2.0107.4kGitHubSite
DSPyStanford NLPProgramming, not prompting, language models.MIT38.9kGitHubSite
Mem0Drop-in persistent memory layer for agents.Apache-2.068.2kGitHubSite
LettaFormerly MemGPTStateful agents with long-term memory that self-improve.Apache-2.025.3kGitHubDocs
STORMStanford OVALResearches a topic and writes a full cited report.MIT31.8kGitHubDemo
H-NetCartesiaNew Oct 31Tokenizer-free, byte-level architecture. A learned dynamic-chunking step replaces BPE tokenization inside the model, and stages can nest (bytes into words into phrases). Cartesia's Oct 28 posts claim a two-stage H-Net matches SmolLM3-3B with about 10x less compute and data; a Hugging Face Transformers PR adding it opened Oct 30.MIT917GitHubPaperBlogModels

Tools and observability 10

ProjectWhat it isLicenseStarsLinks
TrajectoryDatadog LabsNew Nov 4, 2026Observability for AI coding agents: captures sessions from Claude Code, Codex, Cursor and about 20 more via hooks into local timelines, token and cost metrics, and YAML Markers (retries, force-pushes). Datadog LLM Observability export is optional and off by default.Apache-2.042GitHubDocs
Unified Harness Protocol (UHP)HarnessRouterNew Nov 4, 2026StandardOpen draft standard for driving complete agent harnesses (Codex, Claude Code, Hermes, DeepSeek Harness, Gemini CLI) through one HTTP contract: tasks, streaming, sessions, files, cancellation. Shaped like the OpenAI Responses API, with a 104-check conformance suite. HarnessRouter is the self-hosted reference server.Apache-2.04.1kGitHubSpecSite
harness-evalashishpatel26New Nov 4, 2026Did a harness change (AGENTS.md, skills, MCP, hooks) make a coding agent better or worse? Paired, repeated, isolated runs with per-task deltas, bootstrap CI, an A/A noise floor, and CI exit codes. Claude Code, Codex or Ollama.None stated (Caution)3GitHub
agent-desktopVercel LabsNew Nov 2, 2026Rust computer-use for agents via OS accessibility trees, with stable refs and retry-safe actions.Apache-2.08.0kGitHub
OpenBotCopilotKitNew Nov 1, 2026Open-source AI coworkers, each with its own computer (browser, files, tools). Every action is approved before and logged after. Works with any AG-UI agent.MIT9.2kGitHub
Worktrunkmax-sixtyNew Oct 31CLI that makes git worktrees as easy as branches, so you can run many coding agents (Claude Code, Codex) in parallel; adds hooks, LLM commit messages, one-command merge and shared build caches.MIT OR Apache-2.010.0k stars (Nov 2, 2026)GitHubDocs
Lightpandalightpanda-ioNew Oct 31, 2026Headless browser built from scratch in Zig for AI agents; Puppeteer/Playwright via CDP, native MCP server, ~16x less memory than headless Chrome.AGPL-3.0 Caution39.0kGitHubSite
SkyvernAI browser workflow automation.AGPL-3.0 Caution23.3kGitHubSite
AgentOpsAgent monitoring, LLM cost tracking, benchmarking.MIT5.9kGitHubSite
Composio1,000+ toolkits with auth and a sandbox so agents can take action.MIT30.8kGitHubDocs

Agent sandboxes 13

ProjectWhat it isLicenseStarsLinks
Not All Agent Sandboxes Are Created EquallyChristian Posta, kagent.devArticleNew Nov 3, 2026Compares agent sandboxes on four axes: boundary and access controls, environment, lifecycle, and interface and experience. Advice: start with where the agent loop runs. First post in a series, with one follow-up per project.Blog postn/aArticle
Google AXGoogleNew Nov 3, 2026Agent Executor: Google's open agentic orchestration runtime, which runs agent tasks on Agent Substrate.Apache-2.014.5kGitHub
Agent Substrateagent-substrateNew Nov 3, 2026Kubernetes-based core for agent sandboxes. It suspends and wakes idle agents, and egress is denied by default.Apache-2.06.1kGitHub
nononolabs-aiNew Nov 3, 2026Single binary that applies kernel-level restrictions to an agent, with an optional egress proxy.Apache-2.04.8kGitHub
kagentSolo.io / CNCFNew Nov 3, 2026Kubernetes-native agent framework (CNCF sandbox project); publisher of the sandbox series.Apache-2.04.1kGitHubSite
DefenseClawCisco AI DefenseNew Nov 3, 2026Security governance layer for agentic AI, built on NVIDIA OpenShell.Apache-2.0904GitHub
MinimalgominimalNew Nov 3, 2026Runs an agent in a Minimal box: sandboxed sessions built from its own package registry, on your own computer.Apache-2.0237GitHub
Discoboxdiscobox-aiNew Nov 3, 2026Coding-agent sandboxes: a dev box per agent on a copy of the repo (discobox apply).Apache-2.0107GitHub
Substrate envagent-substrateNew Nov 3, 2026Agentic environment abstraction: a shell, file and MCP API on Agent Substrate.Apache-2.071GitHub
Docker SandboxesDockerNew Nov 3, 2026A microVM per coding agent, each with its own Docker daemon.Docker productn/aDocs
OpenShellNVIDIANew Oct 31Policy-governed sandbox runtime for agent fleets. Declarative YAML policies enforced outside the agent: kernel-level file, syscall and network controls, egress proxy, credential brokering, and a formal policy prover. Runs Claude Code, Codex, OpenCode and Copilot CLI on Docker, Podman, microVM or Kubernetes.Apache-2.020.8kGitHubDocs
E2BSecure cloud sandboxes for agents to run code.Apache-2.014.6kGitHubDocs
SPACEPerplexityNew Oct 31Firecracker microVM sandbox behind Perplexity Computer: one VM per agent task, with credentials and network egress controlled outside the VM. E2B-compatible.Apache-2.02.0kGitHubDocsAnnouncement