The model timeline.

Muse Ember (early version)

An early Muse Ember from Meta: free, 1M context, image, audio and video input, blocked in Europe. Our clues pointed to Meta before the reveal.

Iceman Contributor

Free tokens on OpenCode Go with a Go plan, or as Iceman Free in the OpenCode app. Paid for with your prompts. Reasoning, tool calls, 262K context.

Agentic coding · Tool use · 262K context · Adjustable reasoning

Beam (OpenCode preview)

Reflection’s Beam preview, now reached as Beam Free: coding, tool use, 1M context and the released Beam’s benchmark results.

Agentic coding · Tool use · 1M-token context · Adjustable reasoning

Beam (Nous Portal preview)

Reflection’s Beam preview on Nous Portal, now retired: 262K text context, tool calling and always-on reasoning. Beam 1.0 is the way in.

Agentic coding · Agentic reasoning · 262K text context · Tool calling

Glyph Cluster

256K context, function calling and streamed replies. Access through Vercel or Command Code.

256K context · Function calling · Text-only reasoning · Free eligible preview

Fledge Alpha

OpenCode Zen access discontinued. About 61 output tokens/s in our tests, charming but uneven SVG benchmark results, and evidence suggesting a router behind the name.

OpenCode Zen preview discontinued · 1M-token listed context · Text and image input · Adjustable reasoning

Callstack Apex

Revealed as Callstack Apex: a Qwen-based coding model specialized for React Native, React and Next.js.

React Native development · React development · Next.js development · Coding

MiniMax M3.1 Flash (early version)

An early MiniMax M3.1 Flash with expressive SVG benchmark results. Our token counts pointed to MiniMax before the reveal.

SVG benchmarks · 1M-token context · Multimodal input · Adjustable reasoning

Qwen 4

Qwen4-27B, Max, Flash and Plus: release updates, benchmark forecasts and architecture, with a likely 35B-A3B sibling.

Pareto 26.9 (Union of GPT-6 Astra + DeepSeek + GLM)

Pareto 26.9 by Unbiased, first released as Union Alpha. Explore coding benchmarks and SVG results.

Research · Coding · Agentic workflows · Multimodal input

GLM-5.3-Flash

Native multimodal coding, long agent tasks and MIT-licensed model weights.

Coding and agents · Native multimodal input · 1M-token context · MIT-licensed weights

LongCat-2.0

A coding model with native tools and 1M context, revealed as Meituan’s LongCat-2.0.

Agentic coding · Native tool use · 1M-token context · MIT-licensed successor