DeepSeek-V3.2 vs Gemma 4 31B IT

DeepSeek-V3.2

DeepSeek-V3.2 is an MIT-licensed reasoning model trained for parallel tool calls, with thinking chosen per request, for long agent work. It needs about 690 GB of FP8 files, eight 80 GB GPUs in vLLM's recipe and prompts built from DeepSeek's Python. The main risk is neglect, with no file changes since December 2025 and no safety evaluation.

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.

DeepSeek-V3.2

1. Start it with vLLM's recipe, vllm serve deepseek-ai/DeepSeek-V3.2 --tensor-parallel-size 8 --tokenizer-mode deepseek_v32 --enable-auto-tool-choice --tool-call-parser deepseek_v32 --reasoning-parser deepseek_v3
2. It doesn't fit the 107.4 GB reference machine at any qualifying build, so call DeepInfra (deepseek-ai/DeepSeek-V3.2) or SambaNova (DeepSeek-V3.2) instead
3. Outside vLLM, build prompts with encoding/encoding_dsv32.py and wrap parse_message_from_completion_text in error handling, since DeepSeek says it handles well-formed output only
4. Pick thinking or chat with thinking_mode, and read the reasoning from reasoning_content
5. Send developer messages only for search-agent prompts, the only use DeepSeek gives that role
6. Use temperature 1.0 and top_p 0.95 and keep prompts within 128K tokens

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