Ornith-1.5-122B

Ornith-1.5-122B is an open-weight reasoning model in the Ornith family, announced on October 20, 2026. It targets coding and tool-using tasks and uses a mixture-of-experts architecture. The publisher distributes the model through the ornith-ai organization and declares an MIT license.^[1][2]

Architecture and checkpoint

The checkpoint uses the Qwen3_5MoeForConditionalGeneration architecture associated with Qwen3.5. Its configuration includes a language model, a vision encoder and image/video token identifiers. The published chat template accepts text, images and video. These implementation details establish the available input structure; they do not by themselves establish accuracy on visual tasks.^[3][4]

The values above come from the released configuration.^[3] The name's "122B" label is not the exact tensor count: Hugging Face's checkpoint metadata reports 125,086,497,008 BF16 parameters. The weight index records 250,172,994,016 bytes of tensors across 62 Safetensors shards, approximately 250.2 GB before runtime overhead.^[4][5]

Training approach

Ornith describes the 1.5 family's training method as a joint optimization of tasks, scaffolds and solution rollouts. Given an environment or codebase, the system proposes tasks, constructs task-specific instructions and tool orchestration, then attempts solutions. The training rewards cover task validity, difficulty and novelty, the evaluation harness's reliability, and solution success. The family description uses reinforcement learning, including group relative policy optimization, to train these stages together.^[6]

This describes the publisher's training process. It should not be read as a claim that a downloaded checkpoint updates its own weights during ordinary conversations. The August 2026 family article documents the 397B, 35B and 9B releases; their individual benchmark results are not results for the later 122B model.^[6]

Reported evaluations

The following is a selection from the publisher's model-card table checked on October 21, 2026. These are developer-reported results, not an independent replication.^[1]

Terminal-Bench used Claude Code 2.1.126 and averaged five runs. The SWE-bench evaluations used OpenHands, a 256K context window and checks that marked detected reward-hacking solutions incorrect. MCP-Atlas used its 500-task public subset with a ten-minute task timeout. Harnesses, judges and tool access therefore matter when comparing scores.^[1]

The launch post gave Terminal-Bench as 0.813 and HLE as 0.446, without distinguishing tool use for HLE. The model-card table instead lists the values above. They are separate published snapshots, not interchangeable measurements.^[1][2]

Local deployment and formats

The official collection links the BF16 checkpoint and separate FP8, NVFP4, GGUF and MLX repositories. These are different distributions of the same named release, not guarantees of identical numerical behavior or hardware requirements.^[7]

These details are recorded in the original index, variant repository metadata and Ollama's model listing.^{[5][8][9][10][11]} Quantization is relevant to interpreting the publisher's single-device claims. The original BF16 tensor data alone exceeds an 80 GB GPU's memory; a smaller-memory deployment cannot be inferred from the base checkpoint's serving example without a different precision, offloading or another memory configuration.^[5]

For the packaged Ollama build, the documented command is:^[11]

ollama run ornith-1.5:122b

Ollama lists a 77 GB Q4_K_M model and a separate approximately 451-million-parameter BF16 projector. Download size is not a measurement of total working memory, throughput or latency.^[11]

The base checkpoint's chat template separates reasoning text with <think> and </think> delimiters and encodes function calls inside <tool_call> blocks. Applications must use a serving runtime and parsers compatible with that template rather than assuming that raw generated text is already a structured tool invocation.^[4]

Context length and licensing

The configured context limit is 262,144 positions.^[3] The model card suggests YaRN scaling for longer requests and warns that static scaling can reduce quality on shorter inputs. Its roughly one-million-token suggestion is an extension configuration, not the checkpoint's native context setting.^[1] YaRN is a published method for extending rotary-position-embedding context windows; its original research on other models is not an independent long-context evaluation of Ornith-1.5-122B.^[12]

The publisher labels the model MIT-licensed in its announcement and repository metadata. Its project repository carries an MIT license with a 2026 deepreinforce.ai copyright notice. The license permits modification and redistribution subject to retaining the notice and includes a warranty disclaimer.^[2][13] Downloadable weights and a software license should not be confused with a guarantee that an external hosting service is free or has the same terms.

References

1. ^1 ^2 ^3 ^4 ^5Ornith. Ornith-1.5-122B model card. Hugging Face, revision `dfa07bab90e90746ffbfc0b0cba3126eb9265de2`; checked October 21, 2026.
2. ^1 ^2 ^3Ornith. Ornith-1.5-122B launch announcement. October 20, 2026.
3. ^1 ^2 ^3Ornith. Ornith-1.5-122B configuration. Hugging Face, checked October 21, 2026.
4. ^1 ^2 ^3Hugging Face and Ornith. Ornith-1.5-122B repository metadata and chat template. Checked October 21, 2026.
5. ^1 ^2 ^3Ornith. Ornith-1.5-122B Safetensors index. Hugging Face, checked October 21, 2026.
6. ^1 ^2Ornith. Ornith-1.5: From Self-Scaffolding to Self-Improvement. August 2026.
7. ^Ornith. Ornith-1.5 collection. Hugging Face, checked October 21, 2026.
8. ^Ornith. Ornith-1.5-122B-FP8 repository. Hugging Face, checked October 21, 2026.
9. ^Ornith. Ornith-1.5-122B-GGUF files. Hugging Face, checked October 21, 2026.
10. ^Ornith. Ornith-1.5-122B-MLX-4bit repository. Hugging Face, checked October 21, 2026.
11. ^1 ^2 ^3Ollama. ornith-1.5:122b. Checked October 21, 2026.
12. ^Peng, Bowen; Quesnelle, Jeffrey; Fan, Honglu; and Shippole, Enrico. YaRN: Efficient Context Window Extension of Large Language Models. arXiv:2309.00071, 2023; revised February 6, 2026.
13. ^Ornith. MIT license. Project repository, copyright 2026.