DeepReinforce releases Ornith-1.5, claims parity with Claude Opus 4.8

DeepReinforce released Ornith-1.5 on August 19, 2026, an open-weight family of language models spanning three sizes — 397B, 35B, and 9B parameters — that the company says matches or exceeds Anthropic's Claude Opus 4.8 on several coding and reasoning benchmarks. The models ship under the MIT license on Hugging Face, enabling unrestricted commercial use.officechai+1

Benchmarks and Architecture

The flagship Ornith-1.5-397B, a mixture-of-experts model, scores 86.1 on Terminal-Bench 2.1 and 86 on SWE-Bench Verified, compared to Claude Opus 4.8's reported 85 and 85.8 on the same tests, according to DeepReinforce's published results. On DeepSWE, a harder agentic coding benchmark, the model scored 56 — a large jump from 8 on the prior Ornith-1.0 release, though still below Claude Opus 4.8's 59 on the same test. The model trails more clearly on Frontier-Bench v0.1, scoring 13.5 against Claude Opus 4.8's reported 21.1.saascity+2

The architecture extends Ornith-1.0's "self-scaffolding" concept into what DeepReinforce calls a full self-improvement loop: the model proposes its own training tasks, writes task-specific scaffolds, and generates reinforcement-learning rollouts that feed back into training. The 35B MoE variant activates only about 3 billion parameters per token, while the 9B dense model ships with a quantized "Mobile" build compressed to roughly 1.5 GB for on-device use.officechai+2

The Company

DeepReinforce was founded by Dr. Jiwei Li, a Stanford computer science PhD who was named to MIT Technology Review's Innovators Under 35 list. The lab first built its reputation with CUDA-L1 and CUDA-L2, reinforcement-learning frameworks for optimizing CUDA code, before releasing Ornith-1.0 on June 25, 2026. That first release was built on top of Qwen 3.5 Alibaba Group Holding Limited ↗ 5.36% with additional pre-training and post-training.kie+1

Open-Source Context

The release arrives as Chinese and Chinese-linked labs continue narrowing the gap with U.S. frontier models on open weights. Alibaba's Qwen and DeepSeek have made similar claims in recent months, and open-weight releases increasingly benchmark themselves against closed models rather than only against each other. All Ornith-1.5 variants are available on Hugging Face with FP8, GGUF, MLX, and NVFP4 quantized variants, and can be run through Ollama, LM Studio, and AtomicChat.datanorth+1