Checked 23 September 2026. Every figure, date and benchmark score the picture renders is below with the primary source it came from. The video separates four things and so does this file: what was announced, what has been published and measured, what is arithmetic, and what is interpretation.
Where a figure could not be chased to a primary source it is not drawn. Those are listed under “Not put on screen” at the foot.
Primary source: Alibaba Cloud, Apsara announcement, 22 September 2026.
| On screen | Value | Where it comes from |
|---|---|---|
| Qwen 4 status | In training | The announcement |
| The 5 to 10 trillion range | Qwen 4.5 and Qwen 5 | The announcement, which attaches the range to the LATER generations |
| Zhenwu V900 memory | 216 GB | The announcement |
| Zhenwu V900 availability | Planned, Q1 2027 | The announcement |
| Data centre target | Exceeds 20 GW by 2032 | The announcement |
| Self improvement experiment | 33 automated cycles, Artificial Analysis 40 to 45 | The announcement, as a company reported experiment |
What the announcement does not contain, which beat 011 draws as three amber rows because the absence is itself a claim the video makes: no benchmark forecast for any future model, no active parameter count for one, no release date, and no commitment to downloadable ten trillion parameter weights.
Ten trillion is therefore drawn throughout as an ambition attached to later generations, never as the confirmed size of Qwen 4.
Primary sources: Qwen3.8-27B model card, Qwen3.8-Flash-Next model card, Qwen3.8 architecture paper.
Flash Next’s card counts 125B main parameters, 6B active, plus 51B of n gram embeddings and 4B of multi token prediction parameters. Those four figures are what beat 031 stacks.
| Evaluation, as Qwen’s own table reports it | Flash Next | Qwen3.8-27B |
|---|---|---|
| DeepSWE 1.1 | 58.7 | 42.2 |
| SWE-bench Pro | 62.5 | 61.7 |
| Evaluation, as the 27B card reports it | Qwen3.8-27B | Opus 4.6 Max |
|---|---|---|
| SWE-bench Pro | 61.7 | 53.4 |
| Terminal Bench 2.1 | 73.0 | 78.2 |
Two qualifications are drawn on screen rather than kept in this file, because the video states them and the picture must not contradict them:
The 27B card describes a 262,144 token native context, image and video understanding, adjustable thinking effort, Apache 2.0 weights, and quantization links for local use. A maximum advertised context is not a promise of practical context on a 24 GB card, so beat 091 draws a context budget rather than the headline number.
The architecture paper describes the n gram embedding tables being held off the accelerator. That is evidence about a shipping architecture and is drawn as such; it is not a confirmed design for any ten trillion model, which is what beat 034 labels amber.
| Model | Figures drawn | Source |
|---|---|---|
| Nemotron 3.5 Lightning 30B-A3B | 30B total, 3B active, hybrid Mamba and attention MoE, configurable reasoning, released 11 August 2026 | NVIDIA model card |
| Mistral Small 4 | 119B total, image and text input, configurable reasoning, 256K context, Apache 2.0 | Announcement and specifications |
| DeepSeek V4.1 Flash | 552B total, 8B active for input and 16B for output, native visual understanding, MIT weights, released 10 September 2026 | Release notes and weights |
Three points the picture is careful about:
Ideal four bit weight storage is N x 4 / 8 bytes. These are calculations drawn as
calculations, with the caveat on screen: they are not download sizes, not full running
memory, and not a guarantee of quantization quality.
| Parameters | Ideal four bit weights |
|---|---|
| 10 trillion | 5 TB |
| 119 billion | 59.5 GB |
| 27 billion | 13.5 GB |
Real quantized files add overhead and the conversation needs its own memory, per llama.cpp’s quantization documentation and its cache implementation. Offloading changes where weights sit and how fast they are reached; it does not make them stop existing, which is what beats 024 and 025 draw.
RTX 3090 specifications confirm 24 GB of graphics memory. The 64 GB of system memory is an assumed desktop configuration and is drawn as the assumption it is.
DeepSeek R1’s model card documents distilled Qwen and Llama based models. That is the concrete precedent beat 077 shows, and it is a precedent only: it is not a commitment that Alibaba will distribute distilled versions of any future flagship. Beat 081 labels that amber.
These are asserted by the narration and could not be chased to a primary source, so no shot renders them as a figure: