Ornith 1.5 Requirements: VRAM & GPU by Size (9B/35B/397B)

Ornith-1.5 runs across a 8GB to 800GB VRAM range. For on-device testing, the 9B model at Q4 quantization runs from an 8GB-class GPU, while 35B-A3B at Q4 is the practical first choice for solo developers and small teams at 24-32GB-class VRAM.

This table assumes a short-to-medium context, not the full 262,144-token window. If you actually use the full 262K context, the KV cache grows substantially on top of the model weights. For sustained long-context use, budget extra VRAM or combine it with KV cache quantization. Post-quantization sizes above are approximate and vary with implementation, batch size, and context length.

DeepReinforce (Hugging Face org: ornith-ai) released Ornith-1.5 on August 20, 2026; the 35B weights appeared on Hugging Face on August 19-20. It ships under the MIT license. The lineup covers a dense 9B, an MoE 35B-A3B (roughly 35B total parameters, ~3B active), and an MoE 397B, all built via continued pretraining and reinforcement learning on top of Qwen3.5 and Gemma 4. All three sizes support a 262,144-token context, extendable to roughly 1M tokens with YaRN factor 4.0. The 9B also has an Ornith-1.5-9B-Mobile variant for iOS / Android, and 35B-A3B ships with 79 quantized variants compatible with llama.cpp / Ollama / LM Studio.

The core of Ornith-1.5 is its self-improving loop. During training, the model jointly learns to (1) analyze its own solve history and generate harder tasks for itself, (2) build a task-specific scaffold (instructions, tools, decomposition strategy) for that task, and (3) solve the task under that scaffold — with reward from the solution backpropagated through all three stages via GRPO. DeepReinforce calls this a "data flywheel."

The prior generation, Ornith-1.0, optimized scaffolding alongside solution rollouts; Ornith-1.5 extends this to task generation itself, which is the biggest architectural leap between the two generations. Because the model raises its own task difficulty as training proceeds, dependence on hand-designed curricula is reduced.

- Task generation: analyze solve history and author harder tasks
- Scaffold construction: design task-specific instructions, tools, and decomposition strategy
- Solving: solve the task under the constructed scaffold
- Reward from solving backpropagates through all three stages via GRPO

The prior generation, Ornith-1.0, launched on June 26, 2026. The biggest difference at the 35B tier is disclosure: Ornith-1.0's 35B was MoE but never disclosed its active-parameter count, whereas Ornith-1.5 explicitly labels it A3B (~3B active). That makes the "large total, small active" characteristic of 35B-A3B far easier for users to reason about when planning hardware.

DeepReinforce claims 397B's Terminal-Bench 2.1 score of 86.1 slightly edges out Claude Opus 4.8's 85.0, noting it as "the average of 5 independent runs." They also claim 35B-A3B beats the same-size Qwen3.6-35B-A3B across every published coding/agentic benchmark, and even outpaces larger dense models like Gemma 4-31B and Muse Glimmer-30B on agentic coding. All of these are the developer's own measurements — independent third-party verification is not yet established.

- 9B: single-GPU testing and on-device use; the easiest entry point, including the Mobile variant
- 35B-A3B: the go-to for solo developers and small teams — large total parameter count but ~3B active means Q4 quantization stays practical on 24-32GB-class GPUs or a Mac's unified memory
- 397B: datacenter-only, requiring a TP=8 node (e.g., 8x H200 141GB); not a personal-use option

Weight size is roughly "parameter count x bits per weight / 8" (bf16=16 bit, Q8=8 bit, Q4-class formats ~4-5 bit equivalent). For MoE models like 35B-A3B and 397B, remember that the weight footprint that must fit in memory is set by total parameters, while generation speed is set by active parameters — these are easy to conflate. That separation is exactly why 35B-A3B is "large yet fast."

To estimate VRAM for your own GPU, model size, and quantization scheme, plug them into the VRAM Calculator for an estimate and a list of compatible GPUs or Macs.

- Ollama: pull and run with a tag, e.g. ollama run hf.co/<repo>:<quant>
- llama.cpp (GGUF): download a GGUF build from the ornith-ai org on Hugging Face and launch with e.g. llama-server -m ornith-1.5-9b.Q4_K_M.gguf
- vLLM: use vLLM >= 0.19.1 (SGLang >= 0.5.9, Transformers >= 5.8.1) and serve it behind an OpenAI-compatible endpoint

Inference returns chain-of-thought reasoning in a dedicated reasoning_content field, separate from the final answer, which is convenient for applications that don't want reasoning mixed into the response body. Function calling / tool use is supported, so it can be wired into agent pipelines over the OpenAI-compatible API.

For detailed requirements on the Qwen and GLM families, see the Qwen3.8-27B requirements cheat sheet and the GLM-5.2 requirements cheat sheet.

- All benchmark figures in this article are self-reported by developer DeepReinforce; independent third-party verification is not yet established
- Claims of superiority over closed models, including the Claude Opus 4.8 comparison, have drawn some skepticism from the community
- Effective output quality of quantized builds (FP8/INT4/Q4, etc.) needs separate verification from the bf16 baseline
- VRAM figures assume short-to-medium context; long-context use requires additional headroom
- All figures and specs reflect publicly available information as of August 2026