Gemma 4 31B IT vs Swift 1.5 Qwen3.8-Flash-Next

Gemma 4 31B IT

A dense multimodal cousin with documented tool calling, optional thinking and a 256K context, for general agent work and image reading on one 24 GB GPU through Google's 4-bit build. It runs in every major runtime under plain Apache 2.0. The main risk is that Google publishes no safety results and has gone quiet on the repository since July.

Swift 1.5 Qwen3.8-Flash-Next

A large mixture-of-experts cousin, Qwen3.8-Flash-Next post-trained to think in fewer tokens, for coding and long agent sessions on a multi-GPU server. Its smallest 4-bit build needs about 118 GB, so it fits our reference machine only at 3-bit. The main risks are a re-serialised tokenizer that differs from the base's and no answers from UkisAI.

Gemma 4 31B IT

1. On the reference machine run llama-server -hf google/gemma-4-31B-it-qat-q4_0-gguf --hf-file gemma-4-31B_q4_0-it.gguf --jinja -c 32768, adding gemma-4-31B-it-mmproj.gguf for images
2. Or ollama run gemma4:31b for the 20 GB Q4_K_M build
3. In vLLM serve with --enable-auto-tool-choice --tool-call-parser gemma4 --reasoning-parser gemma4, as vLLM's recipe does
4. Thinking is off by default, so pass enable_thinking: true in chat_template_kwargs to turn it on, and expect the reasoning in the thought channel block
5. Drop earlier thoughts from the history except on tool-call turns, as the card asks
6. Use the card's sampling, temperature 1.0, top_p 0.95, top_k 64
7. Validate function names and arguments before running a call, and check JSON output yourself or use the runtime's guided decoding
8. The w4a16 build relies on Marlin INT4 kernels that vLLM marks unsupported on AMD, so use the GGUF on the reference machine

Swift 1.5 Qwen3.8-Flash-Next

1. On a multi-GPU node use vllm serve ukisai/Swift1.5-Qwen3.8-Flash-Next --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder, the parser the base's recipes and the card's SGLang command use
2. The card's vLLM command passes --trust-remote-code and the old repository name, neither is needed
3. On our reference machine only UkisAI's GSQ-RCO IQ3_S GGUF (84.7 GB with the projector) fits, below 4-bit
4. Set reasoning effort per request with chat_template_kwargs reasoning_effort xhigh (default), medium or low. UkisAI's own runs lose about 2.5 points at medium and low
5. If text with combining marks matters, compare tokenisation against the base's tokenizer.json before trusting outputs
6. Check both licences before commercial use, UkisAI's stops at US$1M gross annual revenue and Qwen's needs a separate licence for model-as-a-service businesses