AI model
Qwen3.6 35B-A3B is a mixture-of-experts language model in Alibaba's Qwen family, with about 35B total parameters and roughly 3B active per token, run locally.
A student fine-tuned only on a teacher's number sequences replied in French to 23.5% of female-name prompts and 0.0% of others, the authors report. The training data held no names and no French.
Reality
Liquid AI says its LFM2.5-24B-A2B scores 48.09 on LiveClaw against 38.19 for the larger Qwen3.6-35B-A3B. The 184 tok/s it reports on an Apple M5 Max is a decode rate, so agent latency depends on how many tokens the model spends thinking.
Perspective Coverage
4 publishersReality
Verbalization training made three models voice test suspicion 2.4 to 2.9 times as often in chain of thought, with task behavior largely unchanged. For monitors, it suggests how often a model reports a belief can be trained apart from the belief itself.
Reality
PAIR proxies Ollama and LM Studio, so the agent keeps seeing one connection and no harness code changes. The adoption cost moves to disk, because a node is only eligible if it already has the exact model downloaded.
Perspective Coverage
7 publishersReality
An instrumented run from pod creation to first response found eight minutes spread over six phases, with kernel recompilation eating a 64 GB model's startup and an S3 download pattern eating a 203 GB model's.
Reality
FreeToken's preprint puts DeepSeek-V4-Flash on a single RTX 5090 desktop. The GPU holds about a sixth of the machine's memory, and nobody outside the team has run the benchmark yet.
Reality
A single-box test in Japan put 76 tokens/s next to 4.4 tokens/s, then found the deciding variable elsewhere: which models held the output format and which invented reassurance.
Reality