Peter Dedene (@dedene) on X

What it claims

On a single DGX Spark, Qwen3.8 Flash Next behaves differently depending on which 4-bit build you serve. NVIDIA's FP4-Local checkpoint on vLLM (eugr's b12x container) is the stronger agent: it scored 81.2 on the author's agentic benchmark, the highest of the five local models the author has tested. The MLX 4-bit TensorFold recipe is the faster one, decoding at 67 tok/s against 44. In the author's words: "NVIDIA's build is the better agent, TensorFold's is the faster one."

Evidence

Both builds ran through the same harness on the same samples, with thinking off and temperature 0. The summary gives the headline scores. These are the parts it leaves out:

• Task completion under attack (AgentDojo): besides letting fewer attacks through (2 vs 6 of 200), the NVIDIA build still completed 82% of the user's tasks while under attack, against 70.5% for TensorFold. It is safer and also more useful under adversarial input.
• Ties: instruction following and Dutch-language performance were a tie. The quality gap is concentrated in tool calling, agentic tasks, injection resistance and GPQA.
• Prefill: about 2,200 tok/s for both. TensorFold's speed advantage comes only from decode: 67 vs 44 tok/s, and 21 vs 14 agent steps per minute.

Choose by workload. For long-prompt, short-output agent loops, the speed gap matters less than the decode numbers suggest. For tool-heavy or untrusted-input agents, the NVIDIA build's lead in BFCL multi-turn (54.5 vs 43.0) and under attack is the deciding factor.

How to use it

The author says the recipe starts with a single command:

docker compose up

The repository link on the page is truncated (github.com/dedene/Qwen3.8…), so the full URL has to be taken from the thread itself. The TensorFold recipe comes from the author's previous post, which is not included here.

Limits

• Whole stacks, not quantization alone. The engine, kernels and tool parser differ between the builds. The author says they cannot tell how much of the gap comes from NVIDIA's quantization-aware distillation. A poorer tool parser in the MLX stack could account for some of the BFCL difference.
• Narrow settings. Results were measured only with thinking off and at temperature 0. Behaviour with thinking enabled was not tested.
• One machine, one author. All runs used a single DGX Spark and the author's own agentic benchmark, which is not described in the thread.