DeepSeek Coder V2 16B
Model size, VRAM and Ollama tags

DeepSeek Coder V2 16B hardware requirements

A 12 GB card runs the default tag. A 16 GB card runs q5_K_M or q6_K. A 24 GB card runs q8_0 with the full 128K context. An 8 GB card holds only the 2-bit file.

“VRAM needed” is the file size, plus the key-value cache for a 16K context, plus 1 GB for compute buffers. The cache is small for this model because it uses multi-head latent attention: from config.json, (512 + 64) cached values × 2 bytes × 27 layers = 31,104 bytes per token, which is 0.5 GB at 16K and 4.1 GB at the full 128K. This is a Mixture-of-Experts model, and only 2.4B parameters are active per token, but all 16B have to be loaded, so memory follows the file size.

8 GB cards such as the RTX 3070

The default 8.9 GB tag is larger than the card, so Ollama will place part of the model in system RAM. It runs, but slower than a model that fits. Two alternatives stay entirely on an 8 GB card: the q2_K tag above, at a real cost in output quality, or the older deepseek-coder:6.7b, which is a 3.8 GB download with a 16K context. The Qwen 2.5 Coder 7B page covers a newer model in the same size class.

GGUF sizes are from bartowski/DeepSeek-Coder-V2-Lite-Instruct-GGUF. For the rows with no GGUF file, the Ollama download size is used in the arithmetic. For any other expert count or context length, the MoE VRAM calculator does the same sum.

Ollama tags: which one is the 16B?

Four names point at the same 8.9 GB file: deepseek-coder-v2, deepseek-coder-v2:latest, deepseek-coder-v2:lite and deepseek-coder-v2:16b. The explicit name for it is 16b-lite-instruct-q4_0.

# 8.9 GB, Lite Instruct at q4_0
ollama pull deepseek-coder-v2:16b
ollama run deepseek-coder-v2:16b

# 17 GB, Lite Instruct at q8_0, for a 24 GB card
ollama pull deepseek-coder-v2:16b-lite-instruct-q8_0

• Instruct or base: tags containing lite-instruct are the chat model. Tags containing lite-base are the raw completion weights, at the same sizes.
• Context: Ollama lists 160K for the default tag, which matches the 163,840 positions in the model's config. DeepSeek documents the model as 128K. Ollama loads a much shorter context unless you set one, so raise num_ctx if you need more.
• A tag that does not exist: deepseek-coder:16b-v2. The 16B model is only under deepseek-coder-v2. The separate deepseek-coder library holds the first-generation 1.3b, 6.7b and 33b models.

Source: ollama.com/library/deepseek-coder-v2/tags.

Why does deepseek-coder-v2:16b say it does not support tools?

Because the Ollama build of this model has no tool-calling template. When a client sends a tools array to a model without one, Ollama rejects the request with an error ending in “does not support tools”. The deepseek-coder-v2 library page carries no tools badge, which is how Ollama marks models that accept tool calls. Editor agents and frameworks that rely on Ollama's native tool calling are where most people meet this message.

Three ways forward:

• Use it for what it supports. Chat and code completion work normally. Turn off agent or tool mode in your editor extension.
• Switch to a model with the tools badge. The Ollama tool-calling guide lists them, and Devstral Small 2 is a coding model built for agent use.
• Try a community template. MFDoom/deepseek-coder-v2-tool-calling is an 8.9 GB user upload that carries the tools badge. It is not published by DeepSeek or Ollama and this site has not tested it. DeepSeek reports 0.0 on SWE-Bench for the Lite model, so expect limited results from agent work whichever template you use.

Model specifications

From the model card, its config.json and the paper. The model licence is a custom agreement, so read it before shipping a product on these weights.

DeepSeek Coder V2 benchmarks, as reported by DeepSeek

The 16B Lite model scores 81.1 on HumanEval. The 90.2 figure often quoted for DeepSeek Coder V2 belongs to the 236B model. Both columns below are copied from the table in the DeepSeek-Coder-V2 README for the Instruct models. They are vendor figures and this site has not reproduced them.

Read together, the numbers describe a model that is good at writing a function from a description and weak at multi-step repository work. The full model and its hardware needs are on the DeepSeek Coder V2 236B page.

Is it still worth running?

This section is opinion. DeepSeek Coder V2 dates from June 2024 and newer open coding models have been released since, several of which support tool calling. If you are choosing a model today, start with the Ollama coding-model guide or the best local coding models list. For DeepSeek's current generation see DeepSeek V4.

The reasons to keep the 16B Lite are practical ones: it is an 8.9 GB download that fits a 12 GB card, its key-value cache is small enough that long contexts stay affordable, and the weights allow commercial use. For chat-style help and completion on modest hardware it remains usable. For agents it is the wrong tool.

Frequently asked questions

Sources

• Ollama tag list for deepseek-coder-v2: every tag with its download size
• DeepSeek-Coder-V2-Lite-Instruct model card: parameters, context, licence
• DeepSeek-Coder-V2 on GitHub: benchmark tables and licence section
• DeepSeek-Coder-V2 paper (arXiv:2406.11931)
• bartowski GGUF files: the file sizes in the hardware table