Coding Horizon

Qwen wants 10 trillion. What happens to local AI?

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.

The announcement

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.

Qwen’s published models

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.

The alternatives

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:

The arithmetic

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.

The hardware on the desk

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.

Distillation

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.

Not put on screen

These are asserted by the narration and could not be chased to a primary source, so no shot renders them as a figure: