Kolibri-1 vs Qwen3.5-397B-A17B

Kolibri-1

Kolibri-1 is a German and English reasoning model with a documented tool-call format, switchable thinking and a 262,144-token context, under Apache 2.0. Aleph Alpha's FP8 build needs about 95 GB and runs on vLLM through its plugin, Transformers 5.20 or recent llama.cpp. The main risk is no published safety results or security contact, and few repository answers.

Qwen3.5-397B-A17B

Alibaba's biggest open Qwen3.5, a cousin for hard coding, agent and vision work. Even at 4 bits it needs about 270 GB, so most agents will call it through Alibaba's Model Studio or one of the many providers that host it. The main risk is silence, with no safety results, no agent guidance and template reports without a reply.

Kolibri-1

1. Start it with pip install 'aleph-alpha-inference>=1' then vllm serve Aleph-Alpha/Kolibri-1 --kv-cache-dtype fp8 --reasoning-parser kolibri1 --tool-call-parser kolibri1 --enable-auto-tool-choice
2. Without the kolibri1 tool-call parser and enable-auto-tool-choice, tool calls come back as text in tool_call tags
3. The plugin supports one vLLM minor version, 0.29, so don't upgrade vLLM while it's installed
4. Thinking is on by default at high effort. Pass reasoning_effort none or enable_thinking false in chat_template_kwargs for short tasks
5. Recommended sampling is temperature 1.0, top_p 0.97 and top_k 128
6. Past 262,144 tokens, add --max-model-len 1048576 and the max_position_embeddings override in --hf-overrides, though Aleph Alpha recommends staying at or below 262,144
7. With Transformers 5.20 or later, tokenizer.parse_response returns the answer in content and the reasoning in thinking
8. On the reference machine the FP8 build fits by size, but no document confirms vLLM runs FP8 on gfx1151, and Aleph Alpha ships no GGUF for llama.cpp
9. Validate tool results in your own code, as the card asks

Qwen3.5-397B-A17B

1. The card's server command is vllm serve Qwen/Qwen3.5-397B-A17B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder, and the FP8 build halves the memory
2. It doesn't fit our 128 GB reference machine at any precision, so call a hosted endpoint instead (Model Studio's Qwen3.5-Plus, or a provider serving Qwen/Qwen3.5-397B-A17B)
3. On Model Studio pass "enable_thinking": false directly, while vLLM and SGLang take "chat_template_kwargs": {"enable_thinking": false}
4. Sampling per the card is temperature 0.6, top_p 0.95 and top_k 20 when thinking, and temperature 0.7, top_p 0.8, top_k 20 and presence_penalty 1.5 without
5. For multi-token prediction in vLLM add --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'
6. Keep the context at 128K or more if you rely on thinking, and turn on YaRN only past 262,144 tokens
7. Validate any JSON it returns, since structured output isn't documented