Framework Desktop runs 122B long-context LLMs

This is the kind of local AI benchmark developers actually need: long-context decay curves on real hardware instead of cherry-picked single-point token rates. It strengthens the case for Framework Desktop as one of the most interesting open local-LLM boxes, while also showing that software maturity and context length still dominate the experience once you move past headline specs.

• –Qwen 3.5 35B A3B in Q6_K_L stays relatively strong, posting about 27.8 t/s at 100K context and 19.5 t/s at 250K, which is impressive for a single compact machine.
• –The bigger 122B Qwen 3.5 variants remain technically usable but clearly hit the long-context wall, sliding from roughly 18-21 t/s near 5K context to around 8-10 t/s by 250K.
• –GPT-OSS-20B and GPT-OSS-120B look especially practical on this hardware, suggesting Strix Halo is more than a curiosity for local inference workloads.
• –Community testing around Framework Desktop has already shown backend and ROCm version choices can swing results dramatically, so these numbers are useful as a March 2026 snapshot rather than a final ceiling.
• –Framework's own pitch is that the Desktop can run serious local models on-device; this post shows enthusiasts are already pushing that claim well past Llama-70B-style talking points into 100K+ context experiments.