Ornith-1.5-35B-A3B vs Qwen3-4B-Instruct-2507

Ornith-1.5-35B-A3B

A mid-size coding and terminal-agent cousin with about 3B active parameters, XML tool calls that vLLM and SGLang parse and 262,144 tokens of context. It runs in about 24 GB from Ornith's MLX 4-bit build or 32 GB from its GGUF. The main risk is an absent publisher, with no licence file, safety results or replies to 35 reports.

Qwen3-4B-Instruct-2507

A small, fast cousin for instruction following, short tool-calling jobs and long documents on modest hardware, with no thinking step. It runs in about 3 GB at 4 bits on every common runtime, and two providers host it. The main risk is neglect, with no file changes in over a year, no answers from Alibaba and no safety results.

Ornith-1.5-35B-A3B

1. Start it with ollama run ornith-1.5:35b, or llama-server -hf ornith-ai/Ornith-1.5-35B-A3B-GGUF --port 8000 -c 262144 as the card gives it
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. Use temperature 0.6, top_p 0.95, top_k 20 for general tasks and temperature 1.0 to reproduce the card's benchmarks
6. Pin a revision when you download, since Ornith has replaced files in place, and add mmproj-Ornith-1.5-35B-BF16.gguf (0.9 GB) for image input in llama.cpp
7. Validate JSON output yourself, since structured output isn't documented

Qwen3-4B-Instruct-2507

1. Quickest start is ollama run qwen3:4b-instruct, 2.5 GB. Plain qwen3:4b is the Thinking-2507 edition, not this one
2. On our reference machine use BF16 with vllm serve Qwen/Qwen3-4B-Instruct-2507 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser hermes
3. Use the hermes parser in vLLM or qwen25 in SGLang, the template emits JSON inside tool_call tags
4. Don't pass enable_thinking, this edition never thinks and the card says the flag isn't needed
5. Sampling per the card is temperature 0.7, top_p 0.8, top_k 20 and min_p 0, with presence_penalty up to 2 against repetition
6. Drop the context to 32,768 tokens if memory runs out, as the card suggests
7. Validate JSON yourself or use the runtime's guided decoding, since structured output isn't documented