Unsloth Shrinks Ornith-1.5-122B to Fit on a Single 48GB GPU
Unsloth's dynamic GGUF quants shrink Ornith-1.5-122B down to a 40GB file, letting a frontier-class reasoning and coding MoE fit on a single 80GB GPU.
- Unsloth released dynamic GGUF quants of Ornith-1.5-122B, a 122B MoE reasoning and coding model.
- Base model from Ornith AI is built on Qwen3.5 lineage with self-improvement RL loop.
- Scores 76.9 on SWE-bench Pro and 81.2 on TerminalBench 2.1, beating much larger models.
- Quants range from 40.4 GB (UD-IQ1_M) to 249 GB (BF16); Q4 fits a single 80GB GPU.
- Native 262K context, extendable to ~1M tokens via YaRN RoPE scaling factor 4.0.
- MIT licensed with built-in tool calling; works with vLLM, SGLang, llama.cpp, and Ollama.
Unsloth has published GGUF quantizations of Ornith AI’s Ornith-1.5-122B, a 122-billion-parameter mixture-of-experts reasoning model. The GGUF collection ranges from a 40.4 GB low-bit build to a 249 GB BF16 reference, making the model accessible through llama.cpp-compatible runtimes on high-memory workstations or systems using CPU offload.
The release matters primarily for coding and agent workloads. Ornith reports strong results on SWE-bench Pro, TerminalBench, BrowseComp, and WideSearch, while Unsloth’s quantizations provide several tradeoffs among output quality, memory use, context length, and throughput.
Three training loops move together
Ornith-1.5 follows Ornith-1.0, whose lineage includes Qwen3.5 and Gemma4 plus continued pretraining, mid-training, and post-training. The 122B checkpoint is the smallest announced member of the new family, and Ornith says a mobile-oriented variant is planned.
The new training method jointly optimizes three components that earlier iterations treated more narrowly:
- Task generation: producing new problems for the model to solve instead of depending entirely on a fixed task set.
- Scaffold construction: selecting the prompts, tools, control logic, and intermediate steps used to attempt each task.
- Solution rollouts: improving the model’s actions and answers through reinforcement learning.
This loop lets the training system generate tasks, test different solution strategies, and use the resulting rewards to update the policy. Ornith describes its reward design and evaluation harnesses in the release notes.
Strong coding scores, uneven lead
Ornith’s release reports a large advantage on SWE-bench Pro and competitive agentic scores, while TerminalBench and Humanity’s Last Exam show narrower or negative margins. The benchmarks cover different capabilities: TerminalBench measures command-line work, SWE-bench Pro uses repository-level software issues, BrowseComp tests difficult information retrieval, WideSearch evaluates broad web research, and HLE targets hard knowledge and reasoning tasks.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.