# Atlas with IBM Granite Code 20B (Ollama): More Capacity, Less Window in 2026 > IBM Granite Code 20B (Ollama) is 12GB on disk, Free (self-hosted), and its context window is 8K tokens (8,192). IBM Granite Code 20B (Ollama) is the 20B step in IBM's code line, 12GB on disk and Free (self-hosted). Inside Atlas it is stronger than the 8B on generation quality, but the tradeoff is blunt and worth knowing before you pull: the 20b tag drops to an 8K context, where the 8B instruct advertises 125K. Roughly 14GB to serve lands it on a 16GB card. Pick it when patch quality matters more than session length, and stay at 8B when window matters more than capacity. ## Key takeaways - IBM Granite Code 20B (Ollama) is 12GB on disk, roughly 14GB to serve, and Free (self-hosted). - Its context is 8K tokens (8,192), a hard step down from the 8B instruct tag's 125K. - Instruct and base tags ship at the same 12GB, so it can drive a completion slot or a chat slot. - Agentic tool loops will hit compaction quickly at 8K, so scope tasks tightly. - IBM enterprise licensing and published provenance, the same procurement story as the rest of the Granite line. ## What does IBM Granite Code 20B do better than the 8B? IBM Granite Code 20B (Ollama) buys capacity. The 12GB weights, roughly 14GB to serve, land on a 16GB card and give real capacity gains over the 8B on generation quality. In Atlas that shows up where the model has to hold a nontrivial change in its head and emit a correct patch. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so the practical question about IBM Granite Code 20B is how often that diff is right the first time. The extra parameters help there. IBM ships instruct and base tags at the same 12GB, so you can drive granite-code:20b as a completion engine or a chat model depending on which slot in Atlas you are filling. The Granite line's other constant applies here too: IBM enterprise licensing and published provenance, the same procurement story as the rest of the Granite line, which is often what puts Granite Code 20B on the shortlist ahead of a research release of similar size. ## Why does Granite Code 20B only have an 8K context? The IBM Granite Code 20B (Ollama) tag reports 8K tokens (8,192), a hard step down from the 8B instruct tag's 125K. That is the defining fact about the 20B: it is a capacity upgrade and a context downgrade at the same time, and the two pull in opposite directions for agent work. Register granite-code:20b-instruct in atlas.json with limit.context 8192 and limit.output 4096, matching what the tag actually reports. Do not configure a window the tag does not have. The consequence inside Atlas is that agentic tool loops will hit compaction quickly at 8K, because every turn injects tool output back into the prompt. If window matters more than capacity for your work, stay at 8B, where IBM Granite Code 8B instruct advertises 125K from a 4.6GB download. The 20B is for people whose sessions are short and whose patches must be good, not for people running long build agents across a large repository. ## How do you make an 8K model work in Atlas? Because IBM Granite Code 20B (Ollama) gives you only 8K tokens (8,192), Atlas retrieval has to do the work the window cannot. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, which is what keeps only the relevant chunks in a 2026 prompt this tight. Two Atlas mechanisms carry an 8K model. First, Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so every token spent on context is a whole declaration rather than a truncated fragment. Second, reciprocal rank fusion over semantic and keyword hits ranks those declarations so the few that fit are the few that matter. Generate those vectors with the local Ollama embedder so nothing about the repo is transmitted during indexing, which also means IBM Granite Code 20B and its retrieval layer share the same offline runtime. Used this way, an 8K window is workable for targeted patch generation. It is still not enough for a long autonomous session. ## What hardware does IBM Granite Code 20B need? IBM Granite Code 20B (Ollama) is a 12GB download and needs roughly 14GB of VRAM or unified memory at default context. A 16GB card serves it comfortably. Because the tag caps at 8K tokens (8,192), the KV cache stays small, so 14GB is close to the real total rather than a floor. The small window has one genuine benefit: memory planning for IBM Granite Code 20B is predictable. Long-window models advertise a weight figure and then quietly need many more gigabytes for cache. Granite Code 20B at 8K does not. Plan for roughly 14GB and you are done. Note that Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and parallel subagents share the single served copy of granite-code:20b-instruct, so throughput rather than memory becomes the constraint. Pricing stays Free (self-hosted) regardless of how many subagents you run, which is the compensation for buying a 16GB card. ## When should you pick a different model than IBM Granite Code 20B? Pick the 8B instruct tag instead when session length matters, because IBM Granite Code 20B (Ollama) caps at 8K tokens (8,192) while IBM Granite Code 8B instruct advertises 125K. Atlas lets you switch the active model and provider on the fly with favorites and recents, so keeping both tags is normal. IBM Granite Code 20B earns its 12GB when you want the best Granite patch quality on a 16GB card and your tasks are scoped tightly enough that 8K holds. It is the wrong choice for long agentic loops, for reading large build logs, or for cross-file refactors that require the model to keep many files in mind at once, since agentic tool loops will hit compaction quickly at 8K. Both tags are Free (self-hosted) and both carry IBM enterprise licensing and published provenance, so the decision is purely capacity against window, with nothing else moving. ## Setup 1. Pull it: ollama pull granite-code:20b-instruct (12GB). 2. Plan for roughly 14GB of VRAM or unified memory at default context. 3. Add granite-code:20b-instruct to the ollama provider in atlas.json with limit.context 8192 and limit.output 4096, matching what the tag actually reports. 4. Because 8K is tight, rely on Atlas hybrid retrieval (reciprocal rank fusion over semantic and keyword hits) to keep only the relevant chunks in the prompt. 5. Generate those vectors with the local Ollama embedder so nothing about the repo is transmitted during indexing. ## FAQ ### how to set up atlas with granite code 20b Run ollama pull granite-code:20b-instruct (12GB), plan for roughly 14GB of VRAM, and add granite-code:20b-instruct to the ollama provider in atlas.json with limit.context 8192 and limit.output 4096. ### granite code 20b vs 8b which is better IBM Granite Code 20B is stronger on generation quality, but it caps at 8K tokens (8,192) while the 8B instruct tag advertises 125K. Choose 20B for capacity, 8B for window. ### why is granite-code 20b context only 8k The IBM Granite Code 20B tag reports 8K tokens (8,192). Register it in atlas.json with limit.context 8192 rather than assuming the 8B instruct tag's 125K carries over. ### how much vram does granite code 20b need IBM Granite Code 20B is a 12GB download and needs roughly 14GB of VRAM or unified memory at default context, which lands on a 16GB card. ### is granite code 20b free Yes, IBM Granite Code 20B (Ollama) is Free (self-hosted). It carries IBM enterprise licensing and published provenance, the same procurement story as the rest of the Granite line. ### can granite code 20b handle long agent sessions Not well. With 8K tokens (8,192), agentic tool loops will hit compaction quickly. Lean on Atlas hybrid retrieval, which fuses semantic and keyword hits with reciprocal rank fusion, to keep only relevant chunks in the prompt. ### does granite code 20b have a base tag Yes. IBM ships instruct and base tags at the same 12GB, so you can drive granite-code:20b as a completion engine or a chat model. --- Canonical HTML: https://runatlas.sh/resources/models/granite-code-20b-local Source of truth: aeo_pages row `/resources/models/granite-code-20b-local` (segment: Models) (this file is generated from it, never hand-edited). Licence: Atlas is proprietary with a free core. It is not open source and there is no public source repository.