Volume III · Fusing From the Start

What 2026's Flagships Actually Do

Read four real flagship reports side by side, from late 2023 to late 2025, and the same question keeps recurring underneath the benchmark tables: is this architecture actually early-fused, in the sense Lecture 6 defined, or is it a very well-disguised adapter? The honest answer, for more than one of these labs, is that it is simply not disclosed.

Gemini: "natively multimodal," mechanism withheld

Google DeepMind's Gemini technical report, published in December 2023, and its March 2024 follow-up on Gemini 1.5, describe the model as natively multimodal — pretrained jointly across text, image, audio, and video from the start, rather than built as a text model with vision bolted on afterward, and able to interleave multiple modalities within a single input sequence. That much is a direct, stated claim in the primary source.

What the report does not disclose is the exact tokenization scheme or the specific fusion mechanism underneath that claim — no equivalent of Chameleon's stated token counts, no description of whether images pass through a discrete tokenizer or a continuous encoder. That silence is the boundary of what can honestly be cited from Gemini's own documentation. Any more specific architectural claims about later Gemini versions, 2.x or 3.x, that circulate in blog posts or secondary commentary are not primary sources and should not be cited as fact in either direction.

Qwen2.5-VL and Qwen3-VL: adapter-style, and disclosed as such

The Qwen-VL line sits in useful contrast to Gemini precisely because its technical reports disclose far more. Qwen2.5-VL, published in February 2025, describes a native dynamic-resolution Vision Transformer trained from scratch — using window attention for efficiency, SwiGLU feed-forward layers, and RMSNorm — paired with absolute time encoding for video and dynamic resolution and frame-rate handling that supports hour-long video with second-level event localisation. The 72-billion-parameter variant reportedly matches GPT-4o and Claude 3.5 Sonnet on document and diagram understanding tasks.

Qwen3-VL, published in November 2025, names its innovations precisely rather than leaving them to inference. Interleaved-MRoPE is an upgraded rotary position encoding scheme for jointly modelling spatial and temporal structure across images and video together — a more ambitious cousin of the decoupled-RoPE problem this track's sibling lecture on latent attention worked through for a different reason. DeepStack draws on multiple levels of the Vision Transformer's internal features, not only the final layer, for tighter vision-language alignment. And the model shifts toward explicit text-based timestamp alignment for precise video temporal grounding. It ships with a native 256,000-token context, in both dense variants (2B, 4B, 8B, 32B) and mixture-of-experts variants (30B-A3B, 235B-A22B).

It is worth stating directly, as a point of contrast within this same survey: Qwen2.5-VL and Qwen3-VL are adapter-style architectures — a vision encoder plus a connector feeding an LLM — not early-fusion tokenizer models in the Chameleon or Llama 4 mould from Lecture 6. They belong, architecturally, closer to Volume II than to Volume III, even though they ship in 2025 and outperform much of what either volume produced.

GPT-4o: the objective is confirmed, the mechanism is not

OpenAI's own system card for GPT-4o, together with the accompanying arXiv report, is precise about what it confirms and conspicuously silent about the rest. Confirmed: GPT-4o is trained end-to-end across text, vision, and audio as a single model, with all inputs and outputs processed by the same network, and a disclosed average audio response latency of roughly 232 milliseconds — comparable to human conversational response time. That latency figure is one of the few concrete numbers in the entire report.

Not public: the exact tokenization scheme for images and audio, the specific fusion mechanism — whether early fusion in Lecture 6's sense or some internal adapter architecture — the parameter count, and the training data composition. This is the central teaching point of the whole lecture, stated plainly: OpenAI has confirmed the training objective is unified and end-to-end; the mechanism achieving that unification is not published. Any blog post claiming to know GPT-4o's internal architecture in detail should be treated as speculation, not fact, however confidently it is written.

The spectrum, drawn

Placed side by side, these four reports occupy noticeably different points on an openness spectrum that has nothing to do with model quality. Qwen2.5-VL and Qwen3-VL disclose architecture in the detail a technical report is supposed to carry. GPT-4o discloses an objective and one latency figure. Gemini discloses a philosophy. None of that ranking says anything about which model sees better — it says something about which labs choose to publish mechanism, and which choose to publish results.