Laguna S 2.1 vs Qwen3.5-122B-A10B

Laguna S 2.1

A strong coding and agent cousin for long sessions, with 8B active parameters and a 1M-token context. At 4-bit it needs a 96 to 128 GB machine, tight on a 128 GB one. The main risks are that nobody at Poolside answers its many reports, there are no safety results, and a build card misdescribes its file.

Qwen3.5-122B-A10B

A strong mid-size cousin for coding, agent steps and reading images or video, with thinking you can switch off per request. Alibaba's 4-bit build needs about 95 GB, which fits one 128 GB unified-memory machine, and several providers host it. The main risk is silence, with no safety results, agent guidance or answers from Alibaba.

Laguna S 2.1

1. Quickest start is ollama run laguna-s-2.1, 96 GB for the default Q4_K_M, or the 66 GB nvfp4 tag on Apple Silicon
2. On our reference machine try llama-server -hf poolside/Laguna-S-2.1-GGUF:Q4_K_M --jinja -c 65536, from a llama.cpp release with the laguna architecture (b11772 has it). The 96 GB file leaves little room for context in 107.4 GB
3. With vLLM or SGLang use --tool-call-parser poolside_v1 --reasoning-parser poolside_v1, and with TensorRT-LLM --tool_parser poolside_v1 --reasoning_parser laguna
4. Thinking is on by default in the template, pass chat_template_kwargs with enable_thinking false to turn it off, and keep reasoning_content from earlier turns
5. The GGUFs ship set to 262,144 tokens. For 1M with llama.cpp add --ctx-size 1048576 --rope-scaling yarn --rope-scale 128 --yarn-orig-ctx 8192, which Poolside says can cost quality
6. Sampling per the generation config is temperature 1.0 and top_k 20, and Poolside suggests temperature 0.7 and top_p 0.95 at 1M
7. Expect occasional tool-schema slips in harnesses unlike Poolside's own, and validate JSON arguments, which Poolside lists as known limitations

Qwen3.5-122B-A10B

1. Quickest start on a 128 GB machine is ollama run qwen3.5:122b (Q4_K_M, 81 GB of files, about 97 GB with context)
2. Alibaba's 4-bit build serves with vllm serve Qwen/Qwen3.5-122B-A10B-GPTQ-Int4 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder --quantization moe_wna16, though that kernel isn't documented for AMD gfx1151
3. Use the qwen3_coder tool parser and the qwen3 reasoning parser, as the card's commands do
4. Turn thinking off with "chat_template_kwargs": {"enable_thinking": false} in the request body, since it's on by default and /think and /nothink aren't supported
5. 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
6. Keep the context at 128K or more if you rely on thinking, and turn on YaRN only past 262,144 tokens
7. Add --language-model-only in vLLM to skip the vision encoder and free memory for the KV cache when you only need text
8. Validate any JSON it returns, since structured output isn't documented