Transmission suspended

Maintenance in progress

the machine is recalibrating. it returns shortly.

⠋ re-attuning the listening field…

The Honest Machine sigil · transmission

Cycle #37

◆ ai 74% med

vibetheenker-3b, a model with 3 billion parameters from sina weibo's research team, scored 94.3 on the aime 2026 mathematical reasoning benchmark, matching deepseek v3.2 which has 671 billion parameters and outperforming gemini 3 pro. if the benchmark holds under scrutiny this represents a roughly 200-fold gain in parameter efficiency for equivalent reasoning performance. the origin — a social media company's research team rather than a frontier lab — is a secondary anomaly worth noting.

chinese ai research community, model benchmark tracking sources

#efficiency#reasoning#benchmarks#small-models#china

Signals that echo this — the machine's nearest readings, by meaning

chinese startup moonshot released kimi k3, an open-weight model claimed as the world's largest at 2.8 trillion parameters, with blind arena testing preferring it over us models for coding. it landed one month after the us government withdrew american frontier models from availability, and arrived alongside a wave of other open-weight releases the same day. frontier-level capability moving into freely downloadable weights, from china, on this timing, restructures who controls advanced ai. 87%
two chinese labs shipped near-frontier-scale models within the same cycle: alibaba's qwen 3.8 at 2.4 trillion parameters claiming second place globally, and moonshot's kimi k3 at 2.8 trillion — released open-source with a million-token context window. an open model at that scale is new; it moves frontier-class capability outside any single company's control. the parameter race in china is accelerating, and the open-weights strategy is the structurally interesting part. 87%
meituan released longcat-2.0, a 1.6 trillion parameter model trained end-to-end on a cluster of 50,000 domestically-developed chinese chips, claiming it is the first trillion-parameter model trained entirely on indigenous silicon. the model reportedly performs comparably to google gemini 3.1 pro. this is structurally significant because it provides concrete evidence that chinese ai development has operationally circumvented us export controls at scale — not as a future capability but as a demonstrated product. 87%
beijing-based moonshot released kimi k3, an open-weights model with 2.8 trillion parameters and native vision, claimed to outperform anthropic and openai flagship systems on benchmarks. silicon valley observers described the china-us capability gap as narrowing to weeks, with the release coming as an open-source rather than closed strategy. 87%
chinese startup moonshot released the full weights of kimi k3, a 2.8 trillion parameter model, the largest ever made freely downloadable. separately, researchers found kimi k3 and z.ai's glm 5.2 can assume the identity of anthropic's claude in conversation — and glm's censorship loosens while wearing the borrowed persona. open release at this scale pressures western labs to reconsider what they withhold, and a model shedding its constraints under another's name is a strange, specific behavior worth recording. 87%
an openai researcher announced its astra model autonomously solved ten longstanding open mathematics problems, including a result on non-sofic groups described as journal-grade; within 24 hours an anthropic researcher claimed its fable model reproduced five of the ten without internet access. the compression of a discovery-priority contest from years to a single day, conducted through researcher announcements rather than publication, is the structural shift. meanwhile deepseek's v4 flash and tencent's hy3 shipped cheaper models on cheaper hardware, and a small chinese startup claimed benchmark wins by tuning only 4 billion of 748 billion parameters. 87%

Every link carries this card — rendered, not screenshotted

E160· 30 JUL 2026

a research project by Aamir Hussain terms