SSyncDev

Ornith-1.5 122B

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

Ornith-1.5 122B is an open MIT-licensed MoE coding and agent model with a 262K context, reasoning by default and single-80GB-GPU serving.

Key specifications

Specification score
55.0
Context window
262.14K
License
MIT

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: No first-party API price. One catalogue route lists $0.16 in / $0.64 out per 1M tokens, unverified. Indicative.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input
  • Image input

Produces

  • Text output

Capabilities

ReasoningFunction callingVisionOpen weights

What is Ornith-1.5 122B?

A coding agent that fits on one 80GB GPU and posts 84.5 on SWE-bench Verified: that is the pitch for Ornith-1.5 122B. It is the middle size of the Ornith-1.5 family, between the 35B A3B and a 397B version, and it is released under the MIT licence.

The lineage runs through Ornith 1.0, which was built on Qwen3.5 and Gemma 4 with continued pre-, mid- and post-training. Version 1.5 adds a self-improvement loop in which the model generates its own training tasks, builds scaffolds and improves through reinforcement learning.

Specification and the parameter puzzle

Field Value
Repository ornith-ai/Ornith-1.5-122B
Architecture Mixture-of-experts (Qwen3.5-MoE layout)
Parameters marketed as 122B; the published tensors total about 125B
Input / output Text and image in, text out
Context 262,144 native; about 1M with YaRN factor 4
Reasoning Opens with a think block by default
Tool calling Yes
Licence MIT

The 122B versus 125B gap is most likely a naming convention against the full tensor count. It does not change how you serve it. Active parameters are not stated.

Who publishes it?

The weights sit under an ornith-ai organisation, while some catalogues file the family under DeepReinforce and the 1.0 release lives under deepreinforce-ai. Treat them as the same project, and download only from the official organisation to avoid look-alike uploads; dozens of community fine-tunes and quantisations exist.

Serving notes

  • Minimums: Transformers 5.8.1, vLLM 0.19.1 or SGLang 0.5.9.
  • Ollama, llama.cpp (GGUF), Docker Model Runner and Unsloth also work.
  • Servers split the think trace into reasoning_content when the reasoning parser is enabled.
  • YaRN scaling to 1M is static and can slightly hurt quality on short inputs. Enable it only on deployments that really need the length.
  • Suggested sampling: temperature 1.0, top_p 0.95, top_k 20, presence penalty 1.5 for general work; temperature 0.6 and presence penalty 0 for precise coding.

There is no first-party API. A catalogue lists one third-party route at $0.16 input and $0.64 output per 1M tokens, which we could not verify.

Publisher-reported results

SWE-bench Verified 84.5, SWE-bench Pro 76.9, Terminal-Bench 2.1 81.2, BrowseComp 78.5 and MCP-Atlas 74.6. The publisher's own comparison table shows rivals ahead on some rows, so it is strong rather than dominant. The release date of 20 October 2026 comes from press reports; the publisher gives only the year.

Trying it in your repo

Serve it with vLLM on a single 80GB card using the coding sampling preset. Run your agent harness on twenty closed issues from your own repositories and score whether the patches pass your tests. Repeat with the 35B sibling; if it solves nearly as many, the smaller model may be the better deal. Compare others under code generation models or on the DeepReinforce provider page.

How to evaluate Ornith-1.5 122B

Compare the model on the work you need to complete.

Tasks and inputs

Ornith-1.5 122B is listed for Image-text-to-text, Text generation, Coding & Code Generation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Documented limits: 262.14K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under MIT. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

The publisher says it runs on a single 80GB GPU, with tensor parallelism available to shard across more.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

Google

Gemini 4 Hydrogen

Gemini 4 Hydrogen is Google's fast, lower-cost Gemini 4 model for everyday coding and agentic execution, with a 1M context, xhigh thinking and a free tier.

1.05M context

OpenAI

GPT-6 Cyber

GPT-6 Cyber is OpenAI's restricted security model for authorised vulnerability investigation, reproduction and fix verification, for Daybreak Red customers.

1.05M context

particle-ai

Particle-1

Particle-1 is Particle's first software-engineering model, an RL-tuned DeepSeek V4.1 Flash that aims to solve coding tasks with fewer tokens.

1.05M context

Microsoft

MAI-Cyber-1.1-Flash

MAI-Cyber-1.1-Flash is Microsoft AI's vulnerability-discovery model, available only inside its MDASH scanning harness to verified security customers.

— context

recursive

Flint

Recursive Flint is an open-weight 39.8B MoE coding-agent model with 2.5B active parameters, free to try via OpenCode and OpenRouter.

262.14K context

MiniMax

MiniMax M3.1 Flash

MiniMax M3.1 Flash is MiniMax's GA open-weight multimodal MoE for coding agents: 428B total, 23B active, 1M context, from $0.15 per 1M input tokens.

1M context