Coding Horizon

The Best 16GB Coding Model Is Not The Obvious One

Every figure, size, benchmark and specification the picture puts on screen, chased to a primary source. Vendor pages and model cards first; a specification database only where the vendor does not publish the number itself.

The cards

All four cards named in the video carry 16GB. What differs is the memory type, the interface width and the bandwidth, which is the whole point of that chapter.

Nvidia GeForce RTX 4060 Ti 16GB

Memory: 16GB, GDDR6, 128 bit interface.
Nvidia’s own product page lists the memory size, type and interface width, and does not publish a bandwidth figure for the card.

Source: Nvidia, GeForce RTX 4060 Family specifications — https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4060-4060ti/

Nvidia GeForce RTX 4070 Ti SUPER

Memory: 16GB GDDR6X, 256 bit interface.
Memory bandwidth: 672 GB/s (21 Gbps across 256 bit).

Source: Nvidia, GeForce RTX 40 SUPER Series — https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4070-family/ Corroborated at TechSpot’s specification sheet — https://www.techspot.com/review/2793-nvidia-geforce-rtx-4070-ti-super/

Nvidia GeForce RTX 5080

Memory: 16GB GDDR7, 256 bit interface.
Memory bandwidth: 960 GB/s (30 Gbps across 256 bit). Nvidia’s product page publishes the size, type and interface width but not the bandwidth figure; 30 Gbps modules on a 256 bit bus is where 960 comes from, and it is the figure carried across the specification databases.

Sources: Nvidia, GeForce RTX 5080 — https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5080/ Tom’s Hardware on the 30 Gbps modules and the resulting 960 GB/s — https://www.tomshardware.com/pc-components/gpus/nvidia-rtx-5080-allegedly-adopts-faster-30-gbps-gddr7-modules-delivering-960-gb-s-of-bandwidth-the-remaining-blackwell-lineup-is-expected-to-stick-with-slower-28-gbps-memory

AMD Radeon RX 7800 XT

Memory: 16GB GDDR6, 256 bit interface.
Memory bandwidth: up to 624 GB/s, which is the figure AMD itself publishes.

Source: AMD, Radeon RX 7800 XT product and support pages — https://www.amd.com/en/support/downloads/drivers.html/graphics/radeon-rx/radeon-rx-7000-series/amd-radeon-rx-7800-xt.html

Intel Arc A770 16GB

Memory: 16GB GDDR6, 256 bit interface, 560 GB/s.

Source: Intel, Arc A770 Graphics 16GB specifications — https://www.intel.com/content/www/us/en/products/sku/229151/intel-arc-a770-graphics-16gb/specifications.html

The models

Qwen2.5 Coder 14B Instruct

14.7B total parameters (13.1B non embedding).
Context length 32,768 tokens as configured, with 131,072 available through YaRN scaling. The video quotes the 32k figure, which is the one that applies out of the box.
Ollama’s qwen2.5-coder:14b tag is 9.0GB.

Sources: Qwen model card — https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct Ollama library — https://ollama.com/library/qwen2.5-coder

DeepSeek Coder V2 Lite

16B total parameters with 2.4B active per token, under the DeepSeekMoE framework, so it is a mixture of experts model.
Context length 128k tokens per DeepSeek’s own model card.
Ollama’s deepseek-coder-v2:16b tag is 8.9GB, and the library shows a 160K context window on that tag.

Sources: DeepSeek model card — https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct Ollama library — https://ollama.com/library/deepseek-coder-v2

The video says DeepSeek “reported strong coding results for the time, including HumanEval and MBPP numbers that made much larger models look inefficient” without quoting a figure, so no HumanEval or MBPP number for DeepSeek is put on screen.

StarCoder2 15B

Released by ServiceNow, Hugging Face and Nvidia through the BigCode community. The 15B model specifically was trained by Nvidia; the 3B by ServiceNow and the 7B by Hugging Face.
Nvidia’s release describes the family’s uses as “code completion, advanced code summarization, code snippets retrieval”, and the model card describes training on the fill in the middle objective.
Trained on 600+ programming languages, 16,384 token context window.
Ollama’s starcoder2:15b tag is 9.1GB, which is where “fits sixteen gigabytes comfortably at common four bit quants” comes from.

Sources: Nvidia Newsroom — https://nvidianews.nvidia.com/news/servicenow-hugging-face-nvidia-open-access-llms-generative-ai-enterprise-applications Model card — https://huggingface.co/bigcode/starcoder2-15b Ollama library — https://ollama.com/library/starcoder2

Codestral 22B

22B parameters, trained on 80+ programming languages, 32k context window.
81.1% on HumanEval pass@1 and 78.2% on MBPP, with a leading RepoBench result against the larger code models it was compared with at launch.
Released under the Mistral AI Non-Production License, which is the licensing point the video makes against Apache 2.0.
Ollama’s codestral:22b tag is 13GB with a 32K context window.

Sources: Mistral AI, Codestral announcement — https://mistral.ai/news/codestral Ollama library — https://ollama.com/library/codestral

Qwen3 Coder 30B A3B Instruct

30.5B total parameters with 3.3B activated per token.
Native context 262,144 tokens, extendable to 1M with YaRN.
The model card claims significant performance in agentic coding, browser use and foundational coding, with tool and function calling support and repository scale understanding.
Ollama’s qwen3-coder:30b tag is 19GB, and Ollama describes the model as “the most agentic code model to date in the Qwen series”.

Sources: Qwen model card — https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct Ollama library — https://ollama.com/library/qwen3-coder

Devstral Small 2

24B parameter agentic model for software engineering, built for tool use, codebase exploration and multi file editing.
68.0% on SWE bench Verified (and 55.7% on SWE bench Multilingual).
256k context window on the model card.
Apache 2.0 licensed.
Ollama’s devstral-small-2:24b tag is 15GB, and the library tag shows a 384K context window, which is the “very large context window in the library tag” the video refers to.

Sources: Mistral AI, Devstral 2 announcement — https://mistral.ai/news/devstral-2-vibe-cli/ Ollama library — https://ollama.com/library/devstral-small-2

The benchmarks named

HumanEval — function level Python code generation from a docstring, scored pass@1.
MBPP — Mostly Basic Python Problems, small self contained programming tasks.
RepoBench — repository level code completion, so it rewards using context from across a project rather than a single file.
SWE bench Verified — a human validated subset of real GitHub issues, where the model has to produce a patch that makes the repository’s own tests pass. This is the one closest to what an agent is asked to do, which is why the video weights it as it does.

What is not verified

Memory bandwidth for the RTX 5080 and the RTX 4070 Ti SUPER is not published as a single figure on Nvidia’s own specification pages. Both numbers are derived from the module speed and the interface width Nvidia does publish, and both are the figures carried consistently by the specification databases and launch reviews. They are stated on screen as bandwidth rather than as a vendor quotation.
The 4 bit quantisation claim for StarCoder2 rests on the size of Ollama’s default starcoder2:15b tag rather than on a published quantised file size from BigCode.
Every “feels like”, “is tempting”, “would recommend” judgement in the script is editorial rather than measured, and nothing on screen presents one as a figure.