Coding Horizon

The largest local AI model is bigger than you think

The script arrived finished and was recorded before a shot existed, so this file is not a plan for the narration. It is the list of figures the SHOTS may render, each chased to the page that publishes it.

The asymmetry that makes this step heavier here. The voice has already said every claim below. Where a figure turns out to be wrong the picture can be corrected and the recording cannot, so the rule is: a figure that cannot be sourced is not drawn, and the shot is changed so it does not render the number rather than printing an unsourced figure large.

Checked 21 September 2026. Every figure the script asserts was chased; none of them had to be dropped, and the one that looked wrong on a secondary source turned out to be right on the primary one (see “The one that nearly went wrong” at the foot).


Qwen3.8-27B

Primary: https://huggingface.co/Qwen/Qwen3.8-27B (model card)

figure value where the script says it
parameters 27 billion (27.78B) Start with qwen 3.8 27B. That's twenty seven billion parameters
vision native vision language model, understands images and video it already understands images as well as text
coding and agents stronger autonomous planning, environment feedback, end to end task completion designed for coding and tasks with multiple steps
native context 262,144 tokens, extensible to 1,000,000 not asserted in the script
licence Apache 2.0 not asserted
published 14 August 2026, Alibaba Cloud not asserted

At four bits, 27 billion parameters is 13.5 GB. 27e9 × 0.5 bytes = 13.5 GB. This is the script’s own arithmetic and the script labels it on paper, which is the honest framing: it is a lower bound on the weights and not a download size. Draw it as arithmetic, with the caveat line §13 asks for, never as a measured file size.

LongCat-2.0 (Meituan)

Primary: https://huggingface.co/meituan-longcat/LongCat-2.0 (model card)

figure value where the script says it
total parameters 1.6 trillion a model with 1.6 trillion total parameters
activated per token approximately 48 billion longcat activates about forty eight billion parameters per token
native context 1 million tokens Its million token context support
built for code understanding, repository level edits, agentic workflows built for long context, coding, and tool use
licence MIT not asserted

59 times Qwen. 1.6e12 / 27e9 = 59.3. The script says roughly fifty nine times and then immediately says it does not mean fifty nine times the intelligence, which is the claim the beat is actually making. Draw the ratio; let the correction be the verdict on it.

Kimi K3 (Moonshot AI)

Primary: https://huggingface.co/moonshotai/Kimi-K3 (model card)

figure value where the script says it
total parameters 2.8 trillion (2.78T) It has 2.8 trillion parameters
activated per token 104 billion It activates 104 billion parameters per token
context 1,048,576 tokens supports a million token context
vision native, MoonViT-V2 vision encoder (401M) has native vision
quantization MXFP4 weights, MXFP8 activations, quantization aware training from SFT onward four bit weights with quantization aware training
repository on disk about 1.56 TB The published repository is about 1.56 terabytes
at sixteen bits about 5.6 TB At sixteen bits, it would be 5.6 terabytes
released 16 July 2026, weights under a modified MIT licence not asserted

1.4 terabytes of raw arithmetic. 2.8e12 × 0.5 bytes = 1.4 TB. The script calls it the raw arithmetic ... before overhead and then gives the real published figure of 1.56 TB beside it, which is the correct shape: the arithmetic is the floor and the repository is the fact. Both may be drawn, and they must be drawn as two different kinds of number.

104 times Qwen, and 75 per cent more than LongCat. 2.8e12 / 27e9 = 103.7, and 2.8 / 1.6 = 1.75. Both are the script’s own arithmetic off sourced totals.

Terminal Bench 2.1

From the same model card’s benchmark table:

model score
GPT 5.6 Sol 88.8
Kimi K3 88.3
Claude Fable 5 88.0

The caveat is load bearing and the script already carries it. The model card states that Kimi K3 was evaluated with the Kimi Code harness while the other scores are the best score across harnesses. The script says Those results use different agent setups, so they aren't a controlled comparison of the weights alone, which is the same caveat in the same beat. The picture must carry it too: these three bars are not a controlled comparison and the frame has to say so, per §13.

The serving path

Primary: https://docs.sglang.io/cookbook/autoregressive/Moonshotai/Kimi-K3 (SGLang cookbook)

As of 5 September 2026 the cookbook marks the B300 1×8 Unified Low-Latency and Balanced configurations as verified for a speed round on final weights. Other configurations remain in final verification.

The script says sglang marks its eight b300 GPU configurations for low latency and balanced serving as verified, which matches. It then says Context and concurrency still determine the remaining memory budget, which matches the cookbook’s own framing that the matrix is deployment starting points rather than guaranteed minimum specs.

NVIDIA DGX B300

Primary: https://www.nvidia.com/en-us/data-center/dgx-b300/

figure value where the script says it
GPUs 8 Blackwell Ultra SXM puts eight blackwell ultra GPUs together
total GPU memory 2.1 TB with 2.1 terabytes of total GPU memory
per GPU 288 GB HBM3e not asserted

NVIDIA GeForce RTX 5090

Primary: NVIDIA product page, corroborated by board partner spec sheets (https://www.asus.com/us/motherboards-components/graphics-cards/tuf-gaming/tuf-rtx5090-32g-gaming/techspec/)

figure value where the script says it
VRAM 32 GB GDDR7 An RTX 5090 has thirty two gigabytes of VRAM
memory bandwidth 1,792 GB/s not asserted, but see below

Bandwidth is available and is deliberately not drawn as a figure. Chapter 4’s claim is Memory capacity is how much you can hold. Memory bandwidth is how much you can move each second, which is a distinction rather than a measurement. Printing 1,792 GB/s beside it would answer a question the narration does not ask and would invite a comparison against the B300 that the script never makes.


The hypothetical, and why it is labelled

Take a hypothetical one trillion parameter model is the script’s own word, and the two figures under it are arithmetic on a number that is not a real model:

These are drawn as an equation building term by term with a grey caveat under it, per §10.12 and §13, and the caveat says it is an example rather than a measured model. It must never be drawn as a spec sheet, because there is no product behind it.


The two panels that must stay out of frame

Hugging Face’s Safetensors sidebar widget counts parameters off the tensor files, and it does not agree with either model card’s own prose:

model card says sidebar widget says
Qwen3.8-27B 27 billion 28B params
LongCat-2.0 1.6 trillion total 1.8T params

Neither widget is wrong; it is counting something else, including tensors the headline figure excludes. But the narration says twenty seven billion and 1.6 trillion, so a capture framed to include that sidebar puts a number on screen that contradicts the voice saying a different one, in the same frame, sourced to the same page.

So both captures are cropped to the left hand column, which is what the declared rectangle in each shot does. This is the §8 rule about an annotation contradicting its own picture, arriving through the framing rather than through a caption, and no checker can see it: the asset resolves, the window renders, and the figure is real.

The one that nearly went wrong

A secondary write-up put Claude Fable 5 at 84.6 on Terminal Bench 2.1 rather than the 88.0 the script asserts, which would have made the narration wrong on a figure the picture was about to print large.

Moonshot’s own model card gives 88.0, and the difference is the harness: the card’s own note says non Kimi scores are the best across harnesses, so a single harness number and a best of number are two different measurements of the same model. The script is right, the secondary source was reporting something else, and the resolution is why the figure is taken off the model card rather than off a comparison article.

It is also the reason this video’s bar chart beat carries the harness caveat inside the frame rather than under it.


The photographs and the captures, and where they came from

Section 13.1 makes a photograph house wherever it is EVIDENCE rather than illustration, and names this exact case: a drawn box reading RTX 5090 is a labelled rectangle wearing a product name. Section 13 says the same about a page. Both now carry no ceiling and a high expected frequency, so every beat that rests on somebody’s object or somebody’s page shows it rather than asserting it.

Photographs

beat object file source
037 GeForce RTX 5090 Founders Edition photo/rtx-5090.jpg NVIDIA’s own gallery shot, nvidia.com/en-gb/geforce/graphics-cards/50-series/rtx-5090/, retrieved 22 Sep 2026
050 DGX systems in a data centre aisle photo/dgx-superpod.jpg NVIDIA’s own photograph, nvidia.com/en-us/data-center/dgx-b300/, retrieved 22 Sep 2026
052 NVIDIA DGX B300 photo/dgx-b300.jpg NVIDIA’s own product image, nvidia.com/en-us/data-center/dgx-b300/, retrieved 22 Sep 2026

Each is the maker’s own image of the named thing, not a photograph of something like it, which is section 13.1’s first rule and the same reasoning section 9 applies to marks. Each is credited and dated on screen by the beat’s own source line at the frame edge, because the house has no eyebrow line to put it on, and each carries a short caption naming what it is under its frame.

The 5090 image is the gallery shot rather than the page’s social image. The social image looks down the top edge of the card and shows the fin stack; a viewer who owns one recognises it by its two fans, and the gallery shot is the one with them in it.

None of them is cut out. Section 13.1’s fifth rule was reversed: a photograph keeps its own background and is framed, because a cut out object has been retouched by us and reads as something we drew, which is the opposite of what a photograph is brought in to do. All three happen to be photographed on the maker’s own black, so the frame is doing the work of separating them from the room rather than a mask.

Captures

beat claim it carries file, and the region shown
006 Qwen3.8-27B is 27 billion parameters and multimodal web/qwen-card.png, the model card’s introduction
019 LongCat 2.0 is 1.6 trillion total parameters web/longcat-card.png, the model card’s introduction
070 Kimi K3 is 2.8 trillion parameters web/kimi-card.png, the model card’s introduction
072 context length 1,048,576, and what the model accepts web/kimi-bench.png, the Model Summary rows from Vocabulary Size to Modality
074 88.3 on Terminal-Bench 2.1, against 88.0 and 88.8 web/kimi-bench.png, the Evaluation Results table, four columns, header through Terminal-Bench 2.1
078 the published repository is 1.56 TB web/kimi-files.png, the header and branch rows carrying the size badge
079 sglang marks the eight B300 configurations verified web/sglang-tall.png, the configuration matrix
081 the weight files themselves web/kimi-files.png, five shard rows with their sizes

Every crop is read off the capture’s own pixels and lands on row borders, so no row of a page is cut through the middle. Where a crop shows more than the beat needs — 074 carries the Reasoning and Knowledge rows above the Coding block — that is deliberate: a crop that jumped from the table header to one row would be choosing which benchmarks the viewer sees.

One capture was taken and REJECTED, which is the step section 13 asks for and the reason it asks for it. The model card’s “Evaluation results” sidebar widget was captured for beat 074; it lists Apex Agents, GPQA Diamond, Deep Swe and a long horizon terminal benchmark, and it does not carry a Terminal-Bench 2.1 score at all. A capture that does not show what the beat claims is worse than the drawing it would replace, so the beat uses the model card’s own Evaluation Results TABLE further down the page instead, which does carry it.

Two things the captures corrected rather than merely cited:

What stays drawn. The GPU in beats 045 and 046, the SSD in 047 and the system RAM are CATEGORIES rather than named products, and section 13.1’s test is the actual thing the beat names. An image of a particular vendor’s DIMM standing in for “system RAM” would be the stand in that rule exists to prevent, so those stay drawn.

The three paragraphs added after the first recording

The source script gained three paragraphs on 21 September, after take 3 had already been cut and measured, and the video was re-recorded against the new file. They are at source lines 21, 75 and 115 and became beats 015/016, 053/054 and 081/082.

None of the three carries a figure, which is why this section is short:

Not checked

Everything the script asserts was chased to a primary source. Two things are unverifiable in principle rather than unchecked, and they are opinions the script marks as opinions:

One figure is arithmetic on a hypothetical rather than a measurement, and is labelled as such on screen: