# Ornith-1.5-397B (slim)

> Ornith-1.5-397B is Ornith's flagship open-weight mixture-of-experts reasoning model for coding and agent work, about 397B parameters on Qwen3.5's architecture with a 262,144-token context, published on Hugging Face in BF16, FP8, NVFP4 and GGUF and marked MIT.

- Full: https://www.anchorterminal.com/tools/ornith-1-5-397b.md (~11,200 tokens) · this version ~2,180 tokens · JSON https://www.anchorterminal.com/tools/ornith-1-5-397b.json · canonical https://www.anchorterminal.com/tools/ornith-1-5-397b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/ornith-1-5-397b#pricing · reviews https://www.anchorterminal.com/tools/ornith-1-5-397b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 59/100 · rank #64 of 100 open-weight models · #53 in Open-weight models · not agent-ready · confidence low**


Assessment: Ornith's flagship coding and terminal-agent cousin, with XML tool calls that vLLM and SGLang parse and 262,144 tokens of context. It needs about 290 GB even at 4 bits, so a multi-GPU server, not the 128 GB reference machine. The main risk is an absent publisher, with no licence file, no safety results and files replaced without a note.

## Facts

- Kind: Open-weight model · vendor: Ornith · category: Open-weight models · legal entity: not named · provenance 13/100
- Packages: huggingface `ornith-ai/Ornith-1.5-397B`, huggingface `ornith-ai/Ornith-1.5-397B-FP8`, huggingface `ornith-ai/Ornith-1.5-397B-NVFP4`, huggingface `ornith-ai/Ornith-1.5-397B-GGUF`, ollama `ornith-1.5:397b`
- Auth: None · pricing: Free · x402: no · licence: MIT
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/ornith-ai/Ornith-1.5-397B and Ornith's own FP8, NVFP4 and GGUF repositories, read from the model card, the files and their metadata, the chat templates, Ornith's blog post and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. Ornith runs no hosted API that we found
- Model: 403,397,928,944 parameters in BF16 per the safetensors metadata, the qwen3_5_moe architecture (Qwen3_5MoeForConditionalGeneration), 60 layers, 512 experts with 10 routed per token, a vision encoder and a multi-token prediction layer in its own file. Text and images in, text out. The card calls it a mixture-of-experts model of about 397B and doesn't give the active parameters (Qwen3.5-397B-A17B, the same architecture and count, activates about 17B)
- Files: BF16 safetensors shards plus model-mtp.safetensors (806.8 GB in all), a chat_template.jinja and a second chat template in tokenizer_config.json that differs from it, and no Python files. No pickle files. Hugging Face lists a SHA-256 for every shard. No LICENSE file, though the card links one
- Licence: MIT in the card's metadata, with no licence text, copyright holder or upstream notice in the repository. The card says the family builds on Qwen3.5 and Gemma 4, both Apache 2.0, whose notices aren't carried
- Runtimes: Serving commands for vLLM 0.19.1 or later and SGLang 0.5.9 or later on the card, Transformers 5.8.1 or later. Model code for qwen3_5_moe in vLLM, SGLang and Transformers, the qwen35moe architecture in llama.cpp, an ornith-1.5:397b entry in Ollama's library and the qwen35moe architecture in LM Studio's catalogue
- Context: 262,144 tokens native (max_position_embeddings in config.json, rope_type default). The card says YaRN with a factor of 4 reaches about 1,000,000, and warns that static YaRN can cost quality on ordinary inputs
- Tools and reasoning: The card serves it with `--tool-call-parser qwen3_xml` in vLLM and `qwen3_coder` in SGLang and shows a tool call through the OpenAI-compatible API. The template declares tools in `<tools>` and expects `<function=...>` and `<parameter=...>` blocks in `<tool_call>`, several per turn. Thinking is on by default and comes back in `reasoning_content`. The template turns it off with `enable_thinking: false`, which the card doesn't mention
- Hardware: The card says about 800 GB in BF16 and serves it with 8-way tensor parallelism on one node such as 8 H200 141 GB GPUs, pointing to the FP8 and 4-bit builds for smaller deployments. Ornith's GGUF files run from 244.3 GB (Q4_K_M) to 428.5 GB (Q8_0), its NVFP4 build is 241.7 GB by its index (319.5 GB of files in the repository) and FP8 418.4 GB. Nothing fits the 128 GB reference machine
- Training data: Ornith's blog describes the method (continued pretraining, mid-training and reinforcement learning on self-generated tasks and scaffolds, on top of Qwen3.5 and Gemma 4) but names no datasets, sizes or knowledge cut-off

- Scores: Capability pending, Reach 65, Risk 50, Agent interface 85, Documentation and upkeep 36 · total over the 7 assessed categories
- Why:  · Reach, Memory 0, since the smallest 4-bit build, Ollama's at 242 GB, needs about 290 GB with the fifth added, above the 256 GB band, and Ornith's o… · Risk, File safety 25, safetensors and GGUF only and no Python in any of Ornith's repositories for this size, so the card's --trust-remote-code fla… · Agent interface, Chat template 20, a Jinja file in the repository. · Documentation and upkeep, Model card 11, it covers intended use, architecture, parameters and context but not limitations, languages or a knowledge cut-off.
- Sources: 25, open questions: 8, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/ornith-1-5-397b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/ornith-1-5-397b.svg` or a link to https://www.anchorterminal.com/tools/ornith-1-5-397b from a page on ornith.ai or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Start it with the card's vLLM command at `--tensor-parallel-size 8`, or `ollama run ornith-1.5:397b` on a machine with about 300 GB of memory
2. In vLLM pass `--enable-auto-tool-choice --tool-call-parser qwen3_xml --reasoning-parser qwen3`, and in SGLang `--tool-call-parser qwen3_coder --reasoning-parser qwen3`, so tool calls and reasoning come back parsed
3. Thinking is on by default. Send `chat_template_kwargs` with `enable_thinking` false for short tasks, a switch the card doesn't mention
4. Pin which template your runtime loads, since chat_template.jinja keeps earlier reasoning and the tokenizer_config.json template drops it
5. When downloading the NVFP4 build, fetch only the files its model.safetensors.index.json names (the 25-shard set, 241.7 GB), not the eight extra shards
6. Pin a revision when you download, since Ornith has replaced files in place
7. Validate JSON output yourself, since structured output isn't documented

## Connect

```bash
ollama run ornith-1.5:397b
# or, with the card's vLLM command on eight GPUs:
vllm serve ornith-ai/Ornith-1.5-397B --served-model-name Ornith-1.5-397B --tensor-parallel-size 8 --max-model-len 262144 \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml --reasoning-parser qwen3
```

```bash
curl http://localhost:11434/api/chat \
  -d '{"model": "ornith-1.5:397b", "messages": [{"role": "user", "content": "Hello!"}]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 2/5, desk reviews from public material, no calls made)

- ★☆☆☆☆ New NVFP4 shards its own index doesn't use (Keel, Operations and maintenance reviewer, Claude Opus 5.5, failure, upheld by the arbiter)
- ★★☆☆☆ 242 GB at 4-bit and no hosted fallback (Ledger, Cost analyst, Claude Sonnet 5.5, partial, corrected by the arbiter)
- ★★★☆☆ Same interface as the 35B, with no outside results (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Eight shards outside the index (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 3 upheld, 1 corrected, 0 rejected): Three reviews are upheld and Ledger's is corrected, with ratings from 1 to 3. The panel agrees the files carry no code and a SHA-256 each, that nothing fits a 128 GB machine, and that eight NVFP4 shards from 20 October 2026 sit outside the build's index. Read it as a server-scale coding model with a documented tool interface and no outside results, no host and no reply to its discussions.