The week's most interesting models fit in your pocket

Hear everything, ship nothing

The Speech Event Detector is a 5.8-million-parameter convolutional-Transformer that does one job well: it detects 54 event classes — 20 speaker vocalizations and 34 background sounds. It was trained on 168 hours of real annotated recordings plus synthetic mixtures, and it runs 39 times faster than real time on eight CPU threads. It will not write your essay. It hears the door slam, the laugh, the siren behind the speech — the channel the transcript was never built to capture. For call centers, meetings, and voice agents, the non-verbal channel is becoming structured data, processed on a CPU. m:…6580

The extraction engine

CalOn takes the same philosophy to text. It is a 0.6B-parameter unified encoder for entity, relation, and event extraction that jointly encodes text with type definitions and applies one shared probabilistic formulation across all three tasks, with a probability bias correction to keep confidence estimates honest. The paper reports it outperforming evaluated baselines on most of 40 benchmarks and leading all reported relation extraction benchmarks — processing 120 sentences per second in NER, 5.5 times the throughput of GLiNER2 and 6.7 times GLiFormer. Weights, inference code, and training code are released under Apache 2.0. The extraction layer of the AI stack — the unglamorous work of turning documents into databases — no longer needs a frontier model. It needs a fast one. m:…3238

Decisions without tokens

Intern-Decision and MJev-Spatial push the idea into agency. Intern-Decision is a family of compact multimodal models (0.8B, 2B, 4B) that output structured choices with probabilities via masked language modeling; the 4B version reaches 90.02% unweighted average accuracy across seven evaluation suites, with mean latency around 33–34 ms for a 289-token request on an RTX 4090, and confidence-based handoff lifts accuracy from 73.87% to 87.39% by routing just 42 items to Atria-Dawn-Preview. These are models designed to sit inside an agent loop making thousands of small decisions, where emitting tokens for each one would be a luxury. m:…8766

MJev-Spatial goes further: built on Qwen3.5-4B, it unifies spatial reasoning tasks into a single-pass classification framework that treats each candidate as a class and scores it with the backbone's pretrained query and key projections — generating no answer tokens at all. It runs 5.6 times faster than its base model and 2.5 times faster than the fastest autoregressive baseline, with a 65.64% mean score across 12 benchmarks. When the job is pointing at the right region rather than explaining it, the token stream is pure overhead. m:…3691

The checklist economy

The pattern is a portfolio, not a product: a 5.8M-parameter ear, a 0.6B-parameter reader, sub-4B decision-makers that route to the frontier only when uncertain. Each replaces a frontier-model call for a well-defined job with something cheaper, faster, and auditable — one runs on a CPU, another ships under Apache 2.0. For developers, the decision is to stop asking 'which model' and start asking 'which job': benchmark the narrow task, then buy the smallest model that clears it. The frontier keeps the open-ended work; the checklist goes to the specialists. This suggests the next wave of model value is not in the parameter count but in the calibration — confidence scores you can route on, throughput you can afford, and licenses you can ship. m:…6580 m:…3238