Ornith-1.5-122B Goes Open-Source, Unsloth Ships Quantized Builds

Episode 1 · Unsloth Releases Quantized GGUF of Ornith-1.5-122B (2026-10-20, 2 posts)

Unsloth published GGUF quantized files of the 122B Ornith-1.5 model on Hugging Face, enabling local inference via llama.cpp. Its dynamic UD quants cut size by 82% (e.g. 44.5GB vs 249GB BF16) while retaining about 84% of accuracy.

• Unsloth Releases Ornith-1.5-122B GGUF Quantizations on Hugging Face — LaurentPayot · 2026-10-20
• Unsloth ships 44.5GB quants of Ornith-1.5-122B, 82% smaller with ~84% accuracy — danielhanchen · 2026-10-21

Episode 2 · Ornith releases Ornith-1.5-122B: open-weight coding MoE that runs on a single DGX (2026-10-20, 13 posts)

On October 20–21, the Ornith team released Ornith-1.5-122B, an open-weight 122B-parameter model under the MIT license, positioning it as "strong and fast without trade-offs": the team claims SOTA among same-scale open models while achieving inference efficiency high enough to run on a single NVIDIA DGX card without a large inference cluster. After launch by the original account @ornith (@ornith-ai), the model quickly climbed Hugging Face's trending charts and drew heavy resharing over two days. All benchmark figures so far come from official claims, with no independent reproduction or third-party evaluation yet.

Confirmed

• Ornith-1.5-122B was released with open weights under MIT license, 122B parameters, MoE architecture.
• Per @ornith-ai's model page, it is built on the Qwen3.5 MoE architecture, supports image-text input, text generation and dialogue, ships safetensors weights, and reached Hugging Face trending charts.
• Per @lmoroney, the model was built from Qwen3.5 via additional pretraining, mid-training and post-training; it succeeds Ornith-1.0 and targets coding and agent tasks; SGLang provides day-zero support including reasoning.
• Official benchmarks: SWE-bench Verified 84.5% (0.845), Terminal-Bench 2.1 81.3%, CyberGym 8 (see reshares for details).
• Per @BerryBlushRogers691 relaying the official model card, Ornith-1.5-122B scores 76.9 on SWE-bench Pro vs 43.9 for Qwen3.8-Flash-Next, though it slightly trails that model on Terminal-Bench 2.1.
• Official claims of frontier-level capability runnable on a single NVIDIA DGX; @lmoroney calls it a desktop-runnable coding model.
• Per @Teknium, Nous Portal offered free access to Ornith-1.5-122B for a limited 48 hours.

Unconfirmed

• All benchmark numbers (including SWE-bench Pro 76.9 vs Qwen3.8-Flash-Next) and the "SOTA at same scale" claim come from official promotion or model-card relays; no independent reproduction or third-party evaluation appears in this material.

Why it matters

If the official results hold under independent verification, the model's fully open MIT license and single-machine deployability would sharply lower the barrier to frontier coding/agent capabilities, pressuring both closed APIs and open-weight options that require inference clusters. SGLang's day-zero support means the deployment ecosystem was ready on launch day.

• Open-weights Ornith-1.5-122B launches with 84.5% SWE-bench Verified, runs on a single DGX card — airesearch12 · 2026-10-20
• Ornith-1.5-122B debuts with MIT-licensed open weights and 0.845 SWE-bench Verified — hey_abusiddik · 2026-10-20
• Open-weights Ornith-1.5-122B claims SWE-bench Verified 84.5% running on a single DGX card — alifcoder · 2026-10-21
• Ornith-1.5-122B debuts with MIT weights, claims SWE-bench 84.5% and single-DGX inference — iamfakhrealam · 2026-10-21
• Ornith-1.5-122B open weights debut: 84.5% SWE-bench Verified on a single DGX card — aftahi_ai · 2026-10-21
• Open-weights Ornith-1.5-122B debuts with 84.5% SWE-bench Verified, runs on a single DGX card — _akhaliq · 2026-10-21
• Ornith-1.5-122B open-weights model claims 84.5% SWE-bench Verified, runs on a single DGX card — Teknium · 2026-10-21
• Ornith-1.5-122B: open-source MIT model claims 84.5% SWE-bench on a single DGX card — iamfakhrealam · 2026-10-21
• Ornith-1.5-122B released open-source under MIT, hits 84.5% on SWE-bench on a single DGX card — aftahi_ai · 2026-10-21
• 122B open multimodal model Ornith-1.5, built on Qwen3.5 MoE, trends on Hugging Face — ornith-ai · 2026-10-21
• Ornith-1.5-122B open-weights model lands with 84.5% SWE-bench, runs on a single DGX card — nikola_mr64990 · 2026-10-21
• Ornith-1.5-122B claims 76.9 on SWE-bench Pro vs 43.9 for Qwen3.8-Flash-Next — BerryBlush_Rogers691 · 2026-10-22
• Open-source Ornith-1.5-122B MoE coding model runs on a desk with 86GB memory — lmoroney · 2026-10-22