Unsloth Qwen3.5-35B IQ4_XS tops 100 t/s

This is the kind of local-model result that matters: not leaderboard bragging rights, but a fast enough, obedient enough model that feels usable in real workflows.

• –Roughly 100 t/s on a 3090 changes the experience from “offline batch run” to “interactive assistant,” which is a big deal for coding and agent loops
• –The key signal is persistence under iteration: the model reportedly fixed mistakes and delivered a working Three.js demo after other models kept breaking the app
• –That makes it a strong candidate for agentic tooling like Cline, where multi-step follow-through matters more than one-shot cleverness
• –Unsloth’s own Qwen3.5 docs position the family around 256K context, so this quant sits in a sweet spot of speed, memory efficiency, and long-context practicality
• –The caveat is scope: this is a strong anecdote, not a broad eval suite, so repo-scale refactors and tool-use reliability still need wider testing