Coding Horizon

Your Local Coding Model Is Not Useless, It Is Blind

Every figure this video puts on screen, chased to a primary source.

Checked 13 August 2026.

---

Qwen3 Coder 30B: the local version

30 billion total parameters, about 3.3 billion active per token. The model card gives 30.5B total with 3.3B activated, as a mixture of experts with 128 experts of which 8 are active per forward pass. Source: Qwen — Qwen3-Coder-30B-A3B-Instruct model card https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct

A 256K context window. The same model card gives 262,144 tokens of native context, which is 256K, extensible further with extrapolation methods. Ollama’s library page lists 256K context against both published tags. Source: Qwen — Qwen3-Coder-30B-A3B-Instruct model card https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct Source: Ollama — qwen3-coder https://ollama.com/library/qwen3-coder

Around 19 gigabytes for the common local quantized build. Ollama lists qwen3-coder:30b at 19GB. Source: Ollama — qwen3-coder https://ollama.com/library/qwen3-coder

---

The 480B flagship

480 billion total parameters with 35 billion active. Qwen’s own announcement describes a 480B-parameter mixture of experts model with 35B active parameters, supporting 256K tokens natively and 1M with extrapolation. Source: Qwen — Qwen3-Coder announcement https://qwenlm.github.io/blog/qwen3-coder/

Hundreds of gigabytes, and at least 250GB of memory to run locally. Ollama lists qwen3-coder:480b at 290GB and states that running it locally requires a minimum of 250GB of memory or unified memory. Source: Ollama — qwen3-coder https://ollama.com/library/qwen3-coder

---

Context is paid for in memory

A larger context window needs more memory. Ollama’s context length documentation states directly that setting a larger context length will increase the amount of memory required to run a model. Source: Ollama — Context length https://docs.ollama.com/context-length

The default context Ollama picks depends on available video memory. The documented tiers are 4K context under 24 GiB of VRAM, 32K from 24 to 48 GiB, and 256K at 48 GiB and above. These three figures are the ones the video puts on screen as a stepped chart. Source: Ollama — Context length https://docs.ollama.com/context-length

At least 64,000 tokens for coding tools and agents. The same documentation recommends setting context to at least 64000 tokens for tasks like web search, agents and coding tools, which is what the video shows behind the claim that 32K or 64K already changes how a coding model behaves. Source: Ollama — Context length https://docs.ollama.com/context-length

---

Not checked

These are the video’s judgements rather than published figures, and nothing on screen states them as measurements:

• That 128K is the point where agents stop feeling cramped, and that 256K is the expensive mode to reach for only when the repository and the task justify it. Both are positions the video argues from the memory behaviour above, not published thresholds.
• What the context cache specifically costs in gigabytes at any given window size for this model. It is drawn throughout as an unquantified demand for exactly this reason; the only memory figures rendered as numbers are Ollama’s 19GB and 290GB, its 250GB minimum, its published VRAM tiers, and the machine sizes the video itself names.
• The comparison of local against the best cloud coding models on strength, ease and speed. The video shows this as a shape with no numbers on it, because no benchmark was run for it here.
• The guidance for 32GB, 16GB and 8GB machines. The arithmetic underneath it is sourced, since 19GB of weights against an 8GB machine is a ratio anyone can compute, but the recommendation itself is a judgement.