Qwen3.6-35B-A3B hits 21.7 tok/s on consumer GPUs

// 45d agoBENCHMARK RESULT

Qwen3.6-35B-A3B hits 21.7 tok/s on consumer GPUs

Local LLM benchmarks reveal that Qwen 3.6-35B-A3B, a sparse Mixture-of-Experts model, achieves 21.7 tokens/second on dual RTX 5060 Ti GPUs using hybrid offloading. The model successfully bridges the gap between high parameter counts and consumer hardware, excelling in agentic coding tasks with a 73.4% SWE-bench Verified score.

// ANALYSIS

Sparse MoE architectures are making high-end reasoning viable on consumer-grade setups, though prompt processing remains a significant bottleneck compared to dense models.

–Hybrid offloading (--cpu-moe) provides "free" performance gains by offloading inactive experts to system RAM without sacrificing generation speed.
–The model shows a major reasoning leap, outperforming the Qwen 3.5 dense variant by a substantial margin in agentic benchmarks like Terminal-Bench 2.0.
–PCIe bandwidth limits ingestion efficiency, leaving dense models with a nearly 2x advantage in prompt processing speeds.
–Technical stability remains a challenge; current custom llama.cpp builds crash when combining Gated Delta Net optimizations with hybrid offloading.
–Real-world tests confirm autonomous reliability, with the model successfully completing multi-step tool calls for infrastructure automation.

// TAGS

qwen3.6-35b-a3bllmbenchmarkgpuinferenceopen-weightsreasoningai-coding

DISCOVERED

45d ago

2026-04-17

PUBLISHED

45d ago

2026-04-17

RELEVANCE

9/ 10

AUTHOR

Defilan

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

NEWS22m ago

Developers debate Claude Code and Codex flat-rate pricing

A viral post from DROID (@droidbuilds) sparked a developer debate comparing Anthropic's Claude Code and OpenAI's Codex under a hypothetical $50 monthly flat-rate plan. The discussion highlights the tradeoff between Claude Code's superior reasoning and Codex's deep ecosystem integration when subscription pricing is standardized.

OPEN SOURCE40m ago

Lavish Editor is an open-source, local-first interactive editor designed to streamline human-AI collaboration on HTML artifacts directly in the browser.

Lavish Editor (lavish-axi) is a free and open-source, local-first tool designed to enhance human-AI collaboration on interactive HTML artifacts. Recognizing that AI agents are proficient at generating rich visual and interactive HTML content, Lavish Editor provides a command-line interface (using `npx lavish-axi`) to open these files in a local web browser. Users can select text ranges or pinpoint specific visual elements to leave inline feedback, which can then be read and addressed by the AI agent. Operating entirely locally with zero cloud dependencies, it functions as an Agent Experience Interface (AXI), optimizing token efficiency and human-in-the-loop interactions for complex technical plans, visual designs, and interactive documentation.

UPDATE1h ago

Claude Code drops "workflow" for "ultracode"

Anthropic's developer CLI tool, Claude Code, has released version 2.1.160, replacing "workflow" with "ultracode" as its dynamic workflow trigger. Because "workflow" frequently activated runs unintentionally during regular prompts, the new keyword is designed to provide a cleaner developer experience with fewer false positives.

Qwen3.6-35B-A3B hits 21.7 tok/s on consumer GPUs