Ornith-1.5-35B-A3B is an open-weight reasoning model from Ornith AI, tuned for coding and tool-using agents. It uses the Qwen3.5 mixture-of-experts design: 35 billion parameters, about 3 billion active per token, 262K context. The 5-bit build is 25 GB, so it needs a 32 GB GPU or a large-memory Mac. Pick it for a private coding assistant.
Other good fits
Long inputs, such as a whole repository or a long report, within a 262,144-token window if you have the memory.
+ Pros
Read pros
Pick it when you want a coding assistant that keeps code on your own network. It is trained for tool calls and agent work, and its card gives setups for Ollama, llama.cpp, vLLM, SGLang and agent tools such as Hermes.
Pick it when you need speed on modest hardware. Only about 3 billion parameters work on each token, so it answers much faster than a dense model of the same file size.
Pick it for long material: 262,144 tokens natively. YaRN scaling stretches that to about 1 million, at some quality cost on shorter inputs.
Pick it over plain Qwen3.6-35B-A3B when coding agents are the main job. On Ornith's own benchmark table it scores higher on every coding test, for example 79 against 73.4 on SWE-bench Verified.
We run it ourselves: a 5-bit (Q5_K_M) build is our daily interactive model.
− Cons
Pick something else when you have less than about 24 GB for the model: Phi-4-mini runs on an ordinary laptop.
Pick something else when licence paperwork matters: Qwen3.6-35B-A3B has the same size and speed class and ships an Apache-2.0 licence file.
Pick something else when the hardest tasks matter more than keeping data in-house: a hosted model such as Claude Sonnet handles long, open-ended work better.
The benchmark scores come from Ornith. Try it on a few of your own tasks before you rely on it.
Avoid if
You need a licence text your legal team can read before you ship. The card says MIT, but the repository has no licence file.
Context
The reviewed variant is exactly ornith-ai/Ornith-1.5-35B-A3B; its card states a 262,144-token context window. Optional scaled-context guidance is not the default limit.
Requirements
Ornith's serving example uses two 80 GB GPUs for BF16 weights and long-context headroom. That is publisher guidance, not a measured minimum; runtime, precision, workload and context change memory needs.
Licence & cost
Read licence & cost
Prices checked 23 September 2026: free weights (MIT, as declared on the model card); you pay for the machine.
Build
File size
Fits on
Q4_K_M (4-bit)
21.7 GB
a 24-32 GB GPU with a short context
Q5_K_M (5-bit)
25.3 GB
a 32 GB GPU, or a Mac with 48 GB or more
Q8_0 (8-bit)
37.8 GB
a 48 GB GPU
BF16 (full)
about 70 GB
Ornith's recipe: two 80 GB GPUs for the full 256K context
Long contexts need extra memory on top of the file size. Start at 32K-64K tokens and raise it only when a task needs it.
For vLLM or SGLang you need recent versions (vLLM 0.19.1 or later, SGLang 0.5.9 or later) and the reasoning and tool-call parser flags from the card.
Keep the chat template that ships with the model; tool calls depend on it.
First task
Documentation-based suggestion — not run here.
Read first task
Ask it to map an unfamiliar repository before you change anything.
Install Ollama and run ollama run hf.co/ornith-ai/Ornith-1.5-35B-A3B-GGUF:Q5_K_M (a 25 GB download). With llama.cpp instead: llama-server -hf ornith-ai/Ornith-1.5-35B-A3B-GGUF:Q5_K_M --port 8000 -c 65536.
Point OpenCode at the local OpenAI-compatible endpoint: http://localhost:11434/v1 for Ollama, http://localhost:8000/v1 for llama-server.
Ask: "Read this repository and write a one-page map: what each folder does, where the program starts, and which tests exist."
Check that every file it names exists. Then give it one small change and run the tests yourself.
It is a reasoning model, so each answer starts with a thinking block. Ornith recommends temperature 0.6, top_p 0.95 and top_k 20 for everyday work.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Based on Ornith AI's model card, configuration and GGUF file list on Hugging Face and the Ornith-1.5 blog post, checked 23 September 2026. We did not benchmark it for this review; the team behind this curriculum runs a Q5_K_M build as its daily interactive model.
Ornith-1.5-35B-A3B model card · Sep 23, 2026 — 35B MoE with about 3B active, 262,144-token context, YaRN to about 1M, MIT declared, vLLM/SGLang/Ollama/llama.cpp recipes, recommended sampling, benchmark table vs Qwen3.6-35B-A3B.
Ornith-1.5-35B-A3B config.json · Sep 23, 2026 — Model type qwen3_5_moe (Qwen3.5 mixture-of-experts design).
Ornith-1.5-35B-A3B is an open-weight reasoning model for coding, repository work and agents that call tools. Its Qwen3.5-family mixture-of-experts architecture contains about 35 billion parameters but activates roughly 3 billion per token. Ornith describes training that jointly improves task generation, agent scaffolds and solutions through reinforcement learning. The published context window is 262,144 tokens. It is a candidate for developers who want to run a coding assistant on their own infrastructure: the official card documents vLLM and SGLang servers with OpenAI-compatible APIs, plus GGUF use through Ollama or llama.cpp. Start with a small repository task and check the result against your tests before expanding its role.
Ornith AI · free · self-host
Reasoning output and structured function calls with the documented runtime parsers
Official GGUF distribution for quantized local inference
