Cosmos-Reason2-2B vs Qwen3-30B-A3B-Instruct-2507

Cosmos-Reason2-2B

A small local cousin for reasoning over video and images in robotics, driving and warehouse scenes, returning points, boxes and timed events as JSON. The 4.9 GB file runs in Transformers or vLLM, though NVIDIA asks for 24 GB of GPU memory. Development has moved to Cosmos 3, and the gated licence is revocable.

Qwen3-30B-A3B-Instruct-2507

A quick general-purpose cousin, with 3.3B of 30.5B parameters active per token, for instruction following, tool calling and long documents without a thinking step. It needs about 21 to 23 GB at 4 bits, and several providers host it. The main risk is neglect, with no file changes in over a year, no answers from Alibaba and no safety results.

Cosmos-Reason2-2B

1. Accept the licence on the Hugging Face card with a signed-in account and set HF_TOKEN before downloading
2. Serve it with vllm serve nvidia/Cosmos-Reason2-2B --max-model-len 16384 --media-io-kwargs '{"video": {"num_frames": -1}}' --reasoning-parser qwen3, keeping max-model-len between 8,192 and 16,384 as NVIDIA advises
3. For local video or images, make a folder for them (mkdir -p media) and pass --allowed-local-media-path "$(pwd)/media", so the server can read only that folder, not everything under the directory you start it from
4. Sample video at 4 fps and allow at least 4,096 output tokens so the reasoning isn't cut off
5. For reasoning, add the card's instruction to answer with a think block and then the answer; leave it out for a direct reply
6. Use NVIDIA's prompt files (prompts/*.yaml in nvidia-cosmos/cosmos-reason2) when you want JSON boxes, events or trajectories, and validate the JSON yourself
7. Check its plans before a robot acts on them, since NVIDIA's own safety card warns of collision and manipulation risks

Qwen3-30B-A3B-Instruct-2507

1. Quickest start is ollama run qwen3:30b-instruct, 19 GB. Plain qwen3:30b and qwen3:30b-a3b are the Thinking-2507 edition, not this one
2. On our reference machine use BF16 with vllm serve Qwen/Qwen3-30B-A3B-Instruct-2507 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser hermes, 61 GB of weights fits the GPU share
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. Leave config.json as it is unless you need more than 262,144 tokens, the 1M mode swaps in config_1m.json and runs one sequence at a time
7. Validate JSON yourself or use the runtime's guided decoding, since structured output isn't documented