antirez's DS4 engine: DeepSeek V4 Flash tops local coding on a MacBook Pro

According to a testing report shared on Reddit, DeepSeek V4 Flash running through antirez's custom DS4 engine — using a 2-bit (q2-q4-imatrix) quantization — runs at 25-35 tokens/sec on an M5 Max 128GB MacBook Pro and comes closer to frontier-model quality than any other local setup the author has tried. Even with short, ambiguous prompts, the model reportedly captures intent well and produces usable output on the first pass.

One demonstrative test prompt asked for a single-file pygame script rendering a 3D wireframe terrain with a hand-rolled fBm noise function, three mesh topologies including hexagons, a custom orbit camera, and draggable UI sliders. Local models in the 27-35B range (Qwen3.6, Gemma 4, Phi-4-mini, MiniMax) could only complete the task with heavy scaffolding and still shipped bugs like incorrect depth sorting or missing controls. DeepSeek V4 Flash was the only model that needed no such workarounds — attributed to its 284B-parameter MoE design, where antirez's imatrix recipe keeps attention, router, and shared-expert layers at high precision while aggressively quantizing only the routed experts.

The report also credits the purpose-built engine itself: generic runtimes reportedly can't yet load V4's attention scheme, whereas DS4 pins tool calls to greedy decoding for reliable agentic behavior and uses a disk-based KV cache to avoid re-processing long prompts. The same tester also tried the experimental GLM5.2 model on the same hardware, noting it breaks under long prompts or tool calls. While based on one person's anecdotal testing, the findings offer a concrete signal on how purpose-built engines can make large MoE models genuinely usable on local hardware.