A human
perspective.
An artificial
future.

Anthropic introduced Claude Opus 4 and Claude Sonnet 4, positioning the pair around coding, reasoning, and agent workflows.

The practical question for a developer is how a model behaves across a whole task: reading a repository, making a change, running checks, and responding to feedback. A benchmark is one input, not the entire evaluation.

This is a historical launch example, not a recommendation about today’s model lineup.

Google DeepMind introduced AlphaEvolve, a coding agent that combines language models with automated evaluation and an evolutionary search process.

Its reported applications include data-center scheduling, chip design, and AI training. The interesting pattern is the feedback loop:

1. Propose a candidate.
2. Run a measurable evaluation.
3. Keep useful improvements and iterate.

Ideas are cheap; reliable evaluation is the multiplier.

Google announced Gemini 2.5 Pro with reasoning capabilities and a 1 million-token context window at launch.

A larger window changes what can fit into a request. It does not remove the need to choose relevant material, verify citations, or measure whether the answer actually uses the supplied evidence.

Original explanatory graphic; data from the linked source.

Launch specification from March 2025; not a statement of current limits.

The Gemma 3 family arrived in 1B, 4B, 12B, and 27B parameter sizes. Google described multilingual support across more than 140 languages and context windows up to 128K tokens.

The specifications vary by model: the 1B variant is text-only, while the larger variants add image understanding. Always check the model card for the particular size you intend to run.

Original explanatory graphic; data from the linked source.

Claude 3.7 Sonnet combined ordinary responses and an extended-thinking mode in one model. Anthropic also introduced Claude Code as a command-line coding tool.

For product design, this raises a useful question: when does the user benefit from waiting for a deeper answer, and when is a fast response enough?

Latency belongs in the evaluation alongside answer quality.

Hugging Face’s SmolVLM2 release explored video understanding with models in the 2.2B, 500M, and 256M size range.

That makes a useful counterpoint to the biggest-model headlines: deployment constraints can be part of the research target, rather than something considered after training.

Read the technical details and examples →

Mistral released Small 3, a 24B-parameter model under the Apache 2.0 license, with an emphasis on low-latency language tasks.

• Model quality is one consideration.
• Deployment cost is another.
• License terms are part of the engineering decision too.

Pick the largest model by default.
Start with the task and evaluate what is sufficient.

The Qwen2.5-VL announcement described improvements in document understanding, visual localization, and video comprehension.

These are different capabilities. Reading a chart, locating a button, and understanding a sequence of events should each have their own test cases.

The date shown here follows the publication date displayed on the linked team blog.

DeepSeek announced DeepSeek-R1 and released model weights alongside smaller distilled models.

Access to weights creates room for independent evaluation and deployment experiments. It does not, by itself, settle questions about robustness, training data, or suitability for a particular use case.

The release repository and technical report →

Google introduced Gemini 2.0, starting with an experimental Flash model and demonstrations of tool use and multimodal interaction.

The announcement framed assistants as systems that could do more than answer a single prompt. That makes permissions, observable actions, and recoverable mistakes important parts of the product.

DeepMind’s Genie 2 research demonstrated action-controllable 3D environments generated from a single image prompt.

Its stated purpose was to help train and evaluate embodied agents using diverse simulated environments. A research demonstration is different from a production game engine, but the direction is worth watching.

Anthropic introduced the Model Context Protocol as an open standard for connecting assistants to data sources and tools.

A shared protocol can simplify integration. It does not automatically make a connected tool safe: access control, the scope of each action, and the treatment of untrusted content still matter.

assistant → protocol → tool or data source

Small interface. Large architectural consequences.