GLM-4.7-Flash, Qwen3-TTS opens up, and Claude's new constitution

What shipped

Z.ai (Zhipu)

GLM-4.7-Flash

A lighter GLM-4.7 for local and low-cost use, aimed at coding and agent work.

Liquid AI

LFM2.5-1.2B-Thinking

A reasoning model small enough to run on-device in under a gigabyte of memory.

Alibaba Qwen

Qwen3-TTS

Open text-to-speech with voice design and cloning, an open speech model with voice design and cloning.

Microsoft

VibeVoice

Microsoft's open voice project added an ASR model that handles long recordings with speaker labels.

Anthropic

Claude's new constitution

Anthropic published the document it uses to shape Claude's values and behaviour, rewritten to explain reasons rather than list rules.

A week of small models, voices and one long document about how an AI should think.

Open voice gets good

The most useful release for me was Qwen3-TTS. Alibaba put out an open text-to-speech family with voice design from a text description, voice cloning from a short sample, and multilingual output. Microsoft's VibeVoice project added an ASR model the same week, built for long audio with speakers identified.

Put the two together and you have both halves of a voice pipeline in open weights: speech in, text out, text in, speech out. A year ago that required a paid API for anything that sounded decent. Now I can imagine running it on a machine in my own office.

Voice cloning raises its own problem. The same tool that reads a supplier's quote back to you can fake a voice. I would use it with consent and keep logs.

Small and local

Zhipu released GLM-4.7-Flash, a lighter sibling of the model from December. Liquid AI added LFM2.5-1.2B-Thinking, a reasoning model that fits on a phone in under a gigabyte. Small reasoning models are what will make agents cheap enough to run constantly.

The constitution

On 22 January Anthropic published a new constitution for Claude. The old one read like a list of rules. The new one tries to explain the reasoning behind each value, on the theory that a model which understands why will generalise better than one that only follows a list. It also addresses the question of Claude's own nature more directly than earlier documents did.

Whatever you think of that, publishing it is useful. If you build on a model, you want to know what it was told to care about. I read it the way I would read a supplier's quality manual.

On the constitution, one practical observation. A document like this is also a signal about how the model will behave when instructions run out, which is when business users get surprised. If an agent hits a situation nobody planned for, it falls back on its trained values. Knowing what those are, in plain language, is more useful to me than another benchmark table, and I wish more labs would publish theirs.

What I'd watch next: whether other labs publish their own guiding documents, and whether open voice models get consent features built in.