Agents-A1

generation:

22.0 tokens/s

prompt processing (prefill):

261.0 tokens/s

quant:

ternary

Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.

User reports Millie, a series of compressed agentic models derived from Agents-A1 (a Qwen 3.5 35B-A3B finetune), running on an iPhone 17 Pro with 12 GB RAM. The ternary expert model scores 56% on SWE-bench Verified at 22 t/s decode and 261 t/s pre-fill. The 2-bit expert version scores 60%. The software targets Macs with 16 GB+ memory and Linux gaming PCs with 16 GB+ system RAM and as little as 4 GB VRAM, with a harness forked from OpenAI Codex. The user is seeking testers for AMD GPUs and small 4-8 GB cards.

generation:

97.0 tokens/s

quant:

APEX Compact (GGUF)

Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is shown only when the source reports it; generation speed cannot tell us time to first token. Check the full setup before comparing.

User reports a 35B MoE agentic model that beats Qwen3.6 and DeepSeek V4 Pro in some scenarios. The model is available in the SC117/Agents-A1-Uncensored-MTP-APEX-GGUF repository. A 4B dense variant is also mentioned but is not attractive.