DeepSeekDeepSeek CoderLlamaForCausalLM

Deepseek Coder 33B Instruct — Hardware Requirements & GPU Compatibility

ChatCode

DeepSeek-Coder-33B-Instruct is DeepSeek's 33.3-billion-parameter instruction-tuned code model, initialized from DeepSeek-Coder-33B-Base and further fine-tuned on 2 billion tokens of instruction data for chat-style code generation, completion, and fixing. The underlying Deepseek Coder family was pretrained from scratch on 2 trillion tokens (87% code, 13% natural language in English and Chinese) with project-level context and a fill-in-the-blank training objective, giving it strong project-level completion and infilling ability alongside state-of-the-art open-model results on HumanEval, MBPP, and related coding benchmarks at the time of release. At 33.3 billion parameters, it needs a high-end consumer GPU or a multi-GPU setup once quantized. Context length is 16,384 tokens. Model weights are released under DeepSeek's custom model license, which permits commercial use. It was published in November 2023, alongside 1.3B, 5.7B, and 6.7B siblings and the corresponding base checkpoint.

33.3B params · dense16K contextOtherNov 20234.2K downloads603 likes27.1K quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Decent on a 24 GB GPU

Runs from 16 GB (IQ2_M · 16K)Best on 24 GB (Q4_K_M · ~32 tok/s · 8K)Macs from 18 GB (IQ2_XS · ~10 tok/s)

Other

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run Deepseek Coder 33B Instruct

One recommended way. Every other method, OS and quant is in the step-by-step guide.

Other ways

each opens step-by-step

Run in cloud

Fits on RTX PRO 4500 (32 GB) (7 GB headroom) · Q4_K_M

Generation speed: ~28 tok/s; generation speed
Cost per 1M output tokens: $3.37; per 1M output tokens

Deployon RunPodBest optionfrom $0.34/hr

Compare GPUs →

Run Deepseek Coder 33B Instruct

24.5 GB VRAM at 16K context · needs offload on the RTX 4090 (0.5 GB short)

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run deepseek-coder:33b-instruct-q4_K_M
Official Ollama library build. For the exact Q4_K_M file from TheBloke/deepseek-coder-33B-instruct-GGUF: ollama run hf.co/TheBloke/deepseek-coder-33B-instruct-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and paste https://huggingface.co/TheBloke/deepseek-coder-33B-instruct-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/TheBloke/deepseek-coder-33B-instruct-GGUF@Q4_K_M
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf TheBloke/deepseek-coder-33B-instruct-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/TheBloke/deepseek-coder-33B-instruct-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf TheBloke/deepseek-coder-33B-instruct-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve deepseek-ai/deepseek-coder-33b-instruct --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~73 GB of GPU memory.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model mlx-community/deepseek-coder-33b-instruct
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model mlx-community/deepseek-coder-33b-instruct

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$TheBloke/deepseek-coder-33B-instruct-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
RTX PRO 4500 (32 GB)CheapestCheapest7 GB free	~28 tok/s	$3.37	Deploy on RunPodfrom $0.34/hr (opens in new tab)	32 GB · 7 GB free · ~28 tok/s · $3.37 / 1M tokens
A40 (48 GB)23 GB free	~22 tok/s	$4.48	Deploy on RunPodfrom $0.35/hr (opens in new tab)	48 GB · 23 GB free · ~22 tok/s · $4.48 / 1M tokens
A40 (48 GB)23 GB free	~22 tok/s	$5.15	Deploy on Vast.aifrom $0.40/hr (opens in new tab)	48 GB · 23 GB free · ~22 tok/s · $5.15 / 1M tokens
RTX 5090 (32 GB)7 GB free	~56 tok/s	$3.43	Deploy on RunPodfrom $0.69/hr (opens in new tab)	32 GB · 7 GB free · ~56 tok/s · $3.43 / 1M tokens
RTX 6000 Ada (48 GB)23 GB free	~30 tok/s	$6.80	Deploy on Vast.aifrom $0.73/hr (opens in new tab)	48 GB · 23 GB free · ~30 tok/s · $6.80 / 1M tokens
RTX 6000 Ada (48 GB)23 GB free	~30 tok/s	$6.85	Deploy on RunPodfrom $0.74/hr (opens in new tab)	48 GB · 23 GB free · ~30 tok/s · $6.85 / 1M tokens
RTX 5090 (32 GB)7 GB free	~56 tok/s	$3.79	Deploy on Vast.aifrom $0.76/hr (opens in new tab)	32 GB · 7 GB free · ~56 tok/s · $3.79 / 1M tokens
L40S (48 GB)23 GB free	~27 tok/s	$8.25	Deploy on Vast.aifrom $0.80/hr (opens in new tab)	48 GB · 23 GB free · ~27 tok/s · $8.25 / 1M tokens

Compare GPUs →

How Much VRAM Does Deepseek Coder 33B Instruct Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
10.0 GB	—	—	8.9 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Fits: 16K context · ~66 tok/s
10.8 GB	—	—	9.9 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Fits: 16K context · ~61 tok/s
11.2 GB	—	—	10.5 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Fits: 16K context · ~58 tok/s
12.1 GB	—	—	11.4 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Fits: 16K context · ~54 tok/s
13.7 GB	—	—	12.9 GB	3-bit · i-quant — Importance-weighted 3-bit	Fits: 16K context · ~48 tok/s
14.6 GB	—	—	13.7 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Fits: 16K context · ~45 tok/s
15.0 GB	—	—	14 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Fits: 16K context · ~44 tok/s
15.0 GB	—	—	14.5 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Fits: 16K context · ~44 tok/s
15.4 GB	—	—	14.4 GB	3-bit small — 3-bit small quantization	Fits: 16K context · ~43 tok/s
15.8 GB	—	—	15 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Fits: 16K context · ~41 tok/s
17.1 GB	—	—	16.1 GB	3-bit medium — 3-bit medium quantization	Fits: 16K context · ~38 tok/s
17.5 GB	—	—	18.8 GB	4-bit · legacy — 4-bit legacy quantization	Fits: 16K context · ~38 tok/s
17.9 GB	—	—	17.6 GB	3-bit large — 3-bit large quantization	Fits: 16K context · ~37 tok/s
18.7 GB	—	—	17.9 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Fits: 16K context · ~35 tok/s
19.6 GB	—	—	18.9 GB	4-bit small — 4-bit small quantization	Fits: 16K context · ~34 tok/s
20.8 GB	—	—	19.9 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Fits: 8K context · ~32 tok/s
21.7 GB	—	—	23 GB	5-bit · legacy — 5-bit legacy quantization	Fits: 8K context · ~30 tok/s
23.7 GB	—	—	23 GB	5-bit small — 5-bit small quantization	Fits: 2K context · ~28 tok/s
24.6 GB	—	—	23.5 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Does not fit: needs 28.2 GB
28.3 GB	—	—	27.4 GB	6-bit, very good — 6-bit quantization, very good quality	Does not fit: needs 32 GB
34.2 GB	—	—	35.4 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Does not fit: needs 37.8 GB
67.5 GB	—	—	~66.7 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 71.2 GB

est.calculated, no published file yet

Selected Q4_K_M

4-bit medium quantization

24.5GBat 16K context

Weights: 20.0 GB
KV cache: 4.2 GB
Overhead: 0.3 GB

Fits RTX 4090 (reference) · fits up to 8K context · ~32 tok/s

≥50 tok/s needs ≥1.60 TB/s · RTX 5090 ~56

Which Hardware Can Run Deepseek Coder 33B Instruct?

Q4_K_M · 20.8 GB

Deepseek Coder 33B Instruct (Q4_K_M) requires 20.8 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 28+ GB is recommended. Using the full 16K context window can add up to 3.6 GB, bringing total usage to 24.5 GB. 7 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 100 devices with unified memory can run Deepseek Coder 33B Instruct, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio M1 Max (32 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8–12 GB · RTX 4060, RTX 3060 12GBDoesn't fit — smallest quant needs 13.6 GB8 GB tier →12 GB tier →

16 GB · RTX 5060 Ti 16GBIQ2_M · 15.7 GB · ~24 tok/s · 16K15.7~2416K16 GB tier →

24 GB · RTX 4090referenceQ4_K_M · 24.5 GB · ~32 tok/s · 8Ktight24.5~328K24 GB tier →

Mac 32 GB · Mac Mini M4Q4_K_M · 24.5 GB · ~4 tok/s · 16K24.5~416KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 24.5 GB · ~18 tok/s · 16K24.5~1816KMacs →

Alternatives to Deepseek Coder 33B Instruct

DeepSeek Coder family · 11 models →

Similar models

6tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
MiniCPM V 4.7 35B A3BOpenBMB	—	22.1 GB	~169	same size
Llm Jp 4.1 32B A3b Thinkingllm-jp	—	20.1 GB	~188	same size
Llm Jp 4.1 33B Thinkingllm-jp	—	22.5 GB	~35	same size
Agnes 3.0 FlashAgnes-AI	—	22.1 GB	~35	same size
Ornith 1.5 35B A3Bornith-ai	—	22 GB	~151	same size
NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA	—	19.3 GB	~197	same size

Specifications

Publisher: DeepSeek
Family: DeepSeek Coder
Parameters: 33.3B
Architecture: LlamaForCausalLM
Context Length: 16,384 tokens
Vocabulary Size: 32,256
Release Date: 2023-11-01
License: Other

Resources

Frequently Asked Questions

How much VRAM does Deepseek Coder 33B Instruct need?

Deepseek Coder 33B Instruct requires 20.8 GB of VRAM at Q4_K_M, or 67.5 GB at BF16. Full 16K context adds up to 3.6 GB (24.5 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 33.3B × 4.8 bits ÷ 8 = 20 GB

KV Cache + Overhead ≈ 0.8 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 4.5 GB (at full 16K context)

VRAM usage by quantization

Q4_K_M

20.8 GB

Q4_K_M + full context

24.5 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run Deepseek Coder 33B Instruct?

Yes, at Q5_K_S (23.7 GB) or lower. Higher quantizations like Q5_K_M (24.6 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for Deepseek Coder 33B Instruct?

For Deepseek Coder 33B Instruct, Q4_K_M (20.8 GB) offers the best balance of quality and VRAM usage. Q5_0 (21.7 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 10.0 GB.

VRAM requirement by quantization

IQ2_XXS

10.0 GB~53%

IQ3_XS

14.6 GB~73%

Q4_0

17.5 GB~85%

Q4_K_M ★

20.8 GB~89%

Q5_0

21.7 GB~90%

BF16

67.5 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run Deepseek Coder 33B Instruct on a Mac?

Yes — MacBook Pro 14" M3 Pro (18 GB) and 108 other Macs can run Deepseek Coder 33B Instruct, the smallest at IQ2_XS. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 20.8 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio M1 Max (32 GB).

Can I run Deepseek Coder 33B Instruct locally?

Yes — Deepseek Coder 33B Instruct can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is Deepseek Coder 33B Instruct?

At Q4_K_M, Deepseek Coder 33B Instruct can reach ~230 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~32 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = (bandwidth GB/s ÷ model GB) × efficiency

Example: NVIDIA B200 → 8000 ÷ 20.8 × 0.65 = ~250 tok/s

Estimated speed at Q4_K_M (20.8 GB)

NVIDIA B200

~250 tok/s

NVIDIA GeForce RTX 4090

~32 tok/s

NVIDIA B300

~250 tok/s

AMD Instinct MI350X

~230 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of Deepseek Coder 33B Instruct?

At Q4_K_M, the download is about 20.01 GB. The full-precision BF16 version is 66.69 GB. The smallest option (IQ2_XXS) is 9.17 GB.

Which GPUs can run Deepseek Coder 33B Instruct?

7 consumer GPUs can run Deepseek Coder 33B Instruct at Q4_K_M. Top options include NVIDIA GeForce RTX 5090, AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090. 1 GPU has plenty of headroom for comfortable inference.

Which devices can run Deepseek Coder 33B Instruct?

100 devices with unified memory can run Deepseek Coder 33B Instruct at Q4_K_M (20.8 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.