Open Model WeightsORNITH-AI · FIELD VERIFIED
Primary sources last checked on 24 Oct 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
FRESHNESS & VERIFICATION
A verification check does not pretend the model changed. Open Model Weights updates the verification date after a real primary-source check, while factual-update dates move only when retained source evidence changes.
CONFLICT INTELLIGENCE
+ Save to my shortlistMonitor changes →Local browser monitoring · no background email alerts
DIRECT ANSWER · FEATURED MODEL BRIEFING
Source-grounded editorial layer · reviewed 2026-10-23
Ornith 1.5 122B is an open-weight, Qwen3.5-MoE-derived reasoning and coding model published by Ornith AI with a MIT license declaration and downloadable Safetensors files. Its official model card describes agentic workflows, image-text support and recent vLLM, Transformers and SGLang deployment routes. The 122B family label should not be mistaken for an independently measured GPU-memory requirement; full BF16 weights are far larger than 80GB even if quantized variants can be operated on smaller accelerators. All publisher benchmark and speed tables remain attributed, not verified by Open Model Weights. We classify this model as Notable after a source-pinned weight/license check, but still pending independent comparative quality measurements.
WHY IT MATTERS
A newly released large open sparse model makes externally hosted reasoning capabilities more directly testable in controlled local environments.
The claimed combination of coding, reasoning and multimodal input is worth investigating, with clear separation of actual model artifacts from publisher performance statements.
MODEL-SELECTION VIEW
The official Ornith 1.5 122B repository introduces a sparse mixture-of-experts model intended for reasoning, coding and agentic tasks. It is based on the Qwen3.5 MoE architecture, while the release positions it as a new checkpoint with its own post-training approach and distribution. The rounded model family designation is 122B even though file metadata can count a different number of logical tensor parameters.
A sparse MoE network activates only part of its stored capacity for any given token, but routers, attention, embeddings and expert shards still create substantial memory requirements. A 122B-class checkpoint represented as uncompressed BF16 cannot simply be assumed to fit on an 80 GB accelerator. The publisher's single-GPU claims must be associated with an exact precision variant, runtime and device.
Ornith describes the model as suited to instructions, task decomposition and practical code-agent scenarios. The documented inference interface emits explicit reasoning blocks and separate tool-call structures, which means application developers need to select the supported parser rather than treating every text completion as a safe machine-readable function call.
The card documents compatible recent versions of Transformers, vLLM and SGLang, along with different sampling recommendations for coding and general tasks. When changing runtimes, check that the reasoning parser, end-of-turn markers and call/response turn history remain stable. A successful standalone prompt is not evidence of reliable agent execution over dozens of steps.
The Hugging Face metadata identifies a Qwen3.5 MoE model and displays image-text-to-text support. Operators should distinguish such repository task tags from verified image encoder behavior on their chosen runtime. The exact processor, resizing policy, supported image formats and multimodal prompt template all require a separate smoke test.
Text-only benchmarks are not a proxy for chart reading, screenshot grounding, OCR, spatial questions or tool-use from images. Evaluate images with human-labelled examples, including failures and ambiguous scenes. Track whether the output hallucinates unobserved elements or invents sources, instead of interpreting a single demonstration as complete visual reliability.
Ornith markets the release with favorable tables against similarly sized models and claims an efficient balance of reasoning quality and completion length. The benchmark configurations, answer budgets, context lengths and inference precision may differ between each reported comparison, and some competitor values are developer-supplied.
Open Model Weights has not reproduced the comparative numbers, so this checkpoint is classified Notable rather than Leading. A defensible ranking needs a common test set, public evaluator version, the identical prompt and tool budgets, published exclusions and independent measurements. Until that evidence exists, the table is useful for choosing experiments, not for a universal superiority claim.
The released repository lists downloadable Safetensors weights and a MIT declaration. The publisher also discusses single-accelerator and mobile variants, but those statements must not be automatically assigned to the original full-precision checkpoint. A 122B-class dense-equivalent parameter count at two bytes per stored parameter would be roughly 244 GB before runtime overhead.
Low-bit variants can reduce stored weight memory substantially while altering accuracy, throughput, arithmetic kernels and attention caches. Single-80GB operation depends on an exact quantization and compatible runtime, so assess context length, concurrent batch size, prefill time and peak memory on the target device. Never label a 125B BF16 artifact as a verified 80GB single-GPU model.
Ornith AI declares an MIT license for this checkpoint. Model-license metadata is the starting point, while any base model, training-data restrictions, tokenizer dependencies or runtime libraries must be reviewed separately as applicable. Record the precise repository SHA and weight manifest to make a later audit of file changes possible.
Our assessment is that Ornith 1.5 122B deserves Notable attention for a permissively declared large sparse reasoning model released with operational serving recipes. A practical adoption decision should start with a small controlled evaluation on real coding and reasoning tasks, then expand only if precision-specific quality, tool safety, image processing and hardware costs are acceptable.
IMPLEMENTATION & EVALUATION
Checkpoint-pinned technical evaluation focused on parser behavior, visual evidence, memory and independently repeatable results.
Record model-card SHA, weight manifests, MIT declaration, processing files and exact serving image.
Version-controlled SHA/format checklist and a reproducible text inference sample.
A silently changed tokenizer or runtime shifts benchmark results.
Run structured tool-call conversations and human-labelled image tests with supported parsers.
Reasoning and final output kept separate, valid calls and repeatable media handling.
Unrecognized parser strings yield invalid tools or unreliable visual answers.
Use frozen coding, instruction-following and vision tasks with uniform budgets and baseline models.
Task-level results, failure taxonomy, fixed retries, and transparent scoring.
Publisher benchmark charts omit different evaluation conditions.
Test BF16 and any separately identified low-bit variants on the exact target device.
Measured GPU residency, peak memory, latency, throughput and context-quality tradeoffs.
A 122B model-card label is mistaken for low-memory deployment evidence.
Proposed tests, not OMW benchmark results. Read the official model card at the pinned source revision ↗ Inspect OMW field evidence ↗
PRIMARY & INDEPENDENT SOURCES
FAQ
The publisher maintains a public Hugging Face repository with Safetensors artifacts and an MIT license label. The exact SHA and accessible weight files require independent source-verifier confirmation.
No. The card describes serving variants, but uncompressed BF16 weights for this class are much larger than 80 GB. Any one-device claim must identify precision and runtime.
The publisher documents reasoning-token and tool-call parser conventions. Serving and agent harness software must use the supported formats, and their correctness is not automatic.
The repository advertises image-text support, but the actual visual preprocessing pipeline and per-task image quality require model-specific testing and should not be inferred from the text benchmark chart.
Not according to this site's evidence threshold. It remains Notable until independent, comparable intelligence measurements establish a leadership position for the exact available checkpoint.
The official model repository declares MIT, with separate verification needed for referenced code and dependencies. Do not infer third-party package or base-model rights from the checkpoint label.
MODEL INTELLIGENCE · SOURCE-FIRST
Open Model Weights records Ornith 1.5 122B as a 125.086B-parameter text and image inputs to text, reasoning and code outputs open-weight model in the Ornith 1.5 family, with ornith-ai listed as the developer. The checked source repository provides 63 recognized weight artifacts, format evidence for Safetensors and a verified context window of 262,144 tokens. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares mit. Commercial use allowed under mit conditions. Open Model Weights reports the checked license evidence and classification; this is not legal advice.
No base model is declared in the checked standard lineage metadata, so Open Model Weights does not infer one. Training recipe status: training-described-in-model-card; data disclosure status: not-disclosed-in-standard-metadata-or-obvious-model-card-section.
DEPLOYMENT PROFILE
These bars visualize weight-only decimal-GB estimates. They exclude KV cache, activations, runtime overhead, optimizer state and sharding.
LICENSE & USE
Commercial use allowed under mit conditions
This section summarizes the checked repository evidence. It is not legal advice.
LINEAGE & PROVENANCE
Open Model Weights does not infer a hidden base model or training corpus when the checked sources do not declare one.
RUNTIME SIGNALS
Declared and model-card-mentioned runtimes are discovery signals. They are not treated as Open Model Weights runtime tests unless explicitly measured.
EVIDENCE TRAIL
This is where the article remains auditable: source links on one side, the registry's observed change history on the other.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
01 · WEIGHTS
model-00001-of-00062.safetensorsmodel-00002-of-00062.safetensorsmodel-00003-of-00062.safetensorsmodel-00004-of-00062.safetensorsmodel-00005-of-00062.safetensorsmodel-00006-of-00062.safetensorsmodel-00007-of-00062.safetensorsmodel-00008-of-00062.safetensorsmodel-00009-of-00062.safetensorsmodel-00010-of-00062.safetensorsmodel-00011-of-00062.safetensorsmodel-00012-of-00062.safetensorsmodel-00013-of-00062.safetensorsmodel-00014-of-00062.safetensorsmodel-00015-of-00062.safetensorsmodel-00016-of-00062.safetensorsmodel-00017-of-00062.safetensorsmodel-00018-of-00062.safetensorsmodel-00019-of-00062.safetensorsmodel-00020-of-00062.safetensorsmodel-00021-of-00062.safetensorsmodel-00022-of-00062.safetensorsmodel-00023-of-00062.safetensorsmodel-00024-of-00062.safetensorsmodel-00025-of-00062.safetensorsmodel-00026-of-00062.safetensorsmodel-00027-of-00062.safetensorsmodel-00028-of-00062.safetensorsmodel-00029-of-00062.safetensorsmodel-00030-of-00062.safetensorsmodel-00031-of-00062.safetensorsmodel-00032-of-00062.safetensorsmodel-00033-of-00062.safetensorsmodel-00034-of-00062.safetensorsmodel-00035-of-00062.safetensorsmodel-00036-of-00062.safetensorsmodel-00037-of-00062.safetensorsmodel-00038-of-00062.safetensorsmodel-00039-of-00062.safetensorsmodel-00040-of-00062.safetensorsmodel-00041-of-00062.safetensorsmodel-00042-of-00062.safetensorsmodel-00043-of-00062.safetensorsmodel-00044-of-00062.safetensorsmodel-00045-of-00062.safetensorsmodel-00046-of-00062.safetensorsmodel-00047-of-00062.safetensorsmodel-00048-of-00062.safetensorsmodel-00049-of-00062.safetensorsmodel-00050-of-00062.safetensorsmodel-00051-of-00062.safetensorsmodel-00052-of-00062.safetensorsmodel-00053-of-00062.safetensorsmodel-00054-of-00062.safetensorsmodel-00055-of-00062.safetensorsmodel-00056-of-00062.safetensorsmodel-00057-of-00062.safetensorsmodel-00058-of-00062.safetensorsmodel-00059-of-00062.safetensorsmodel-00060-of-00062.safetensorsmodel-00061-of-00062.safetensorsmodel-00062-of-00062.safetensorsmodel.safetensors.index.json02 · LICENSE
Classification describes the checked license terms; it is not legal advice.
03 · HARDWARE
Weight-only decimal-GB estimate from total parameter count; excludes KV cache, activations, runtime overhead, optimizer state and sharding.
04 · FORMATS
“Not observed” means no matching repository artifact/name/dtype signal was found; it does not prove no third-party conversion exists.
05 · LINEAGE
06 · TRAINING ASSETS
07 · RUNTIME SUPPORT
Open Model Weights separates model-repository evidence from external runtime-release evidence. An external support statement does not change the official model repository's weight formats and does not imply an Open Model Weights inference benchmark.
08 · SOURCES
09 · VERIFICATION HISTORY
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
MODELS FOR AI AGENTS · ROLE EVIDENCE
Reasoning model. A candidate for a reasoning or coordination role. No end-to-end agent performance is inferred from its model category.
The curated notable tier is not an agent-readiness rating. Model-specific tool-call success, long-horizon task accuracy and compatible runtime configurations are not independently established here.
Source-linked license: mit (allowed commercial-use classification; see original terms).
Context: 262144 configured tokens.
BF16 weights: 250.17 estimated decimal GB, not runtime VRAM.
Compare agent model candidates → · Agent evaluation method → · Inspect original model repository ↗
These records are separately sourced. They do not replace independent model-plus-runtime tests.
No model-specific agent capability evaluation is recorded for this checkpoint. Missing evidence is unknown, not a failed task.
Read evidence details and benchmarks → · Use the agent model selector →