Agents
nanocoder
Community-built terminal coding agent: bring your own model, keep code local.
★ 2.5k starsCustom
View guide →Beam — Reflection's open 501B MoE model built for agentic coding: 23B active parameters, a 1M-token context, Apache-2.0 weights, and Mirror, its open-source terminal harness
Beam is an open-source ai agents project by reflection-oss (reflection-oss/mirror on GitHub). It has 6 GitHub stars and 0 forks, is released under the Apache-2.0 license, and was last updated on GitHub on 16 October 2026.
Download a checkpoint from Hugging Face in the numerics your hardware suits: the main BF16 repo, or the FP8 and NVFP4 variants published alongside it — all Apache-2.0.
Run it with the day-0 vLLM, SGLang or transformers support. Plan for datacenter GPUs: the practical self-host route is the NVFP4 checkpoint on four B200s, which leaves room for long-context KV cache and concurrent requests.
Install Reflection's open-source harness, launch it inside your project and sign in: Mirror reads and edits files, runs commands, and works parallel tasks in separate repository worktrees.
Built for coding agents and tool use: 81.6% SWE-bench Verified, 80.1% Terminal Bench v2.1 and 80.8% MCP Atlas on the model card (BF16), with a dedicated tool-calling grammar.
Hybrid sliding-window and global attention grows the KV cache by roughly 36 KiB per token — about 36 GiB at the full 1M-token context, roughly a sixth of the naive rate.
The model card reports comparable solution quality to leading open models using roughly 3x fewer inference-time reasoning tokens; effort is adjustable (low, medium, high, xhigh, max) and thinking can persist across turns.
Apache-2.0 weights in three precisions, a native transformers implementation and published SFT/RL recipes — a frontier-scale model with no API lock-in.
The launch recipe serves the BF16 checkpoint across 16 GPUs behind an OpenAI-compatible endpoint.
vllm serve reflection/Beam-1.0-501B-A23B --tensor-parallel-size 16 --served-model-name beamThe native implementation lands in transformers 5.20.1+; the tokenizer ships a response template that parses reasoning, answers and tool calls for you.
pip install -U "transformers>=5.20.1"
from transformers import pipeline
pipe = pipeline(
"text-generation",
model="reflection/Beam-1.0-501B-A23B",
device_map="auto",
)Reflection's terminal coding agent: the package is reflection-mirror, the command is mirror.
uv tool install reflection-mirror
cd /path/to/your/project
mirrorReflection (Reflection AI), a US lab. The weights are published on Hugging Face under the reflection org; the code home is the reflection-oss GitHub org.
Yes — the model weights and the Mirror harness are both Apache-2.0. Running a 501B model still needs serious GPU hardware, or a hosted inference provider.
Datacenter GPUs. The BF16 checkpoint is about 1 TB; the practical self-host route is NVFP4 (about 280 GB) on Blackwell hardware — the launch guide plans for four B200s. FP8 suits Hopper/Ampere setups.
Against API-only frontier models, Beam is downloadable, self-hostable and Apache-2.0. Against other open MoE models, its card leads with agentic results (81.6% SWE-bench Verified, 80.1% Terminal Bench v2.1), and it ships its own open-source harness, Mirror.
Download a checkpoint from Hugging Face, serve it with vllm serve (or SGLang/transformers), and install Mirror with uv tool install reflection-mirror to put the model to work in your repo.
Community-built terminal coding agent: bring your own model, keep code local.
ZCode — Z.ai's AI coding workspace: desktop app, browser interface and terminal agent in one open-source harness
Up to 20x faster AI coding with smart model routing and full self-hosting.
The official collection of reference Model Context Protocol servers.
Compression gateway that cuts AI coding agents' token bills up to 85%.
The official TypeScript SDK for Model Context Protocol servers and clients.
The most-starred other AI Agents projects in the LibreAI catalogue — the realistic shortlist if Beam is not the right fit.
A personal AI that really does things — on any OS, with your data staying yours.
An agentic skills framework & software development methodology that works.
Agent harness optimization system with skills, memory and security for Claude Code, Codex and more
The agent that grows with you — a Python framework for adaptive AI agents.
DeepSeek's plugin-based agent harness where everything is a plugin.
Free and open — grab the source, star the repo, and start creating.