AI model
Qwen3.5 is Alibaba's Qwen family of open-weight language models, released in dense and mixture-of-experts variants from sub-1B to larger sizes.
NVIDIA and Princeton's Dao AI Lab cache SSM inputs instead of state, targeting a decode step where SSM layers are the largest component up to 100K tokens. ReplaySSM is already integrated into TensorRT-LLM, vLLM and FlashInfer, NVIDIA says.
Reality
One developer's self-hosted Langfuse traced 91 of 1,387 agent model calls in 14 days because only one of ten profiles had its keys. The plugin fails open by design, so nine untraced profiles, one a coding agent with more than five times the traced calls, raised no error.
Reality
Researchers writing on LessWrong found only Nemotron 3 Super and Qwen3 32B changed refusals when test suspicion was removed from their reasoning traces. In most other models, test talk in a trace looks like general caution, so it is weak evidence of eval gaming.
Reality
One developer's audit of Claude Code found 97 of 200 agent steps could run on a local model, against 4 of 100 whole requests. That makes the agent step the unit to route on, on evidence from one person's sessions and one RTX 4070.
Reality
Jared Palmer's Kev now ships as 0.8B, 4B and 9B variants that score typed answer options over frozen Qwen3.5 weights, and on his own development comparison hosted Jev still scores higher than the largest of them.
Reality
NVIDIA's SWE-Serve scores the same 627 patches twice on 19 SGLang tasks, once with the live-serving tests and once without. The pass rate falls from 69.4% to 45.9%, and 242 of the 276 live tests came from SGLang itself.
Reality
Bespoke Nimble's LoRA fine-tune of Qwen3.5-9B scored 90% against Jev's 93% on an eval it curated itself, and five more replications landed at scales between 421M and 35B parameters. No standard benchmark exists for the category yet.
Reality
Per-tensor layout maps now drive his GGUF releases. The sensitivity data behind them came out of more than 1,000 quantizations of two Qwen3.5 models, scored by KL divergence against bf16 on wikitext-2-raw.
Reality
Publishing weights and publishing something a team can actually deploy are different acts, and N2.5 lands on both sides of that line depending on which tier you pick and whether its weights exist yet.
Reality
p-for-llm puts a 29-expert, top-1-routed mixture of experts on an ESP32-P4 and got two points on Hacker News. The routing, not the quantization, is the load-bearing part.
Reality