Unreal Agent: an open coding harness 39% cheaper than Codex

It hit Codex's Terminal-Bench 4.0 score at $1,428 vs $2,350 by firing tools async. Plus: AMD's 14 open hybrid checkpoints cut KV-cache up to 98%.

Nowline SEP 23 2:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Same score as Codex, 39% less spend

    Unreal Labs open-sourced Unreal Agent, a coding harness that matched Codex's 57.9% pass on Terminal-Bench 4.0 (with GPT-6 Astra) for $1,428 versus Codex's $2,350. It fires tool calls asynchronously — logging an 'in-progress' event and reasoning further instead of blocking on each result. Ships as a Go library, a standalone runner, and a Harbor-compatible benchmark runner.

  • AMD turns Llama and Qwen into long-context hybrids

    AMD released 14 open Zebra-HyLo checkpoints that 'upcycle' Llama-3.2-1B/3B and Qwen3-1.7B into MLA + Mamba-2 or Gated-DeltaNet hybrids, cutting KV-cache 92-98% and extending context to 64K (up to 2M on eight MI300X). Apache-2.0 on Hugging Face, with training code on GitHub — long context without pretraining from scratch.

  • Build this weekend: 64K context on a small VRAM budget

    Grab a Zebra-HyLo 1-3B checkpoint and you get 64K-token context with a KV-cache up to 98% smaller than the base Transformer — the memory wall that usually caps local long-context. Point it at a big codebase or a long-document pipeline that wouldn't fit before.