Stable DiffCoder 8B Instruct
Stable-DiffCoder-8B-Instruct is a code-focused mask diffusion language model by ByteDance-Seed, built on the Seed-Coder architecture and fine-tuned from Stable-DiffCoder-8B-Base. It uses block diffusion continual pretraining followed by supervised fine-tuning and is instruction-tuned for alignment with user intent. The model card lists an 8K context length, text-generation pipeline, transformers library support, MIT license, and safetensors format.
At a glance
- Parameters
- 8.25B
- Architecture
- llama
- Context length
- 8,192
- License
- mit
- Disk space
- 15.4 GB
- Software
- Transformers
Will it run on your machine?
Save your machine to see a personalized rating and its reasoning.
Add my hardwareBest for
- Code generation and completion
- Structured code editing
- Code reasoning tasks
- Instruction-following for programming prompts
- Handling of low-resource programming languages
Tradeoffs
- The source does not specify hardware requirements for local use.
- No performance benchmarks or evaluation results are provided in the supplied facts.
- Compatibility is only stated with transformers v5.3.0.
- Licensing details are limited to the stated MIT license.
- No testing or validation information is provided.
Ways to run it
Transformers · See official guide
Repository-specific command in publisher documentation; confirm dependencies and hardware in the source.
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
device = 'cuda'
model = AutoModelForCausalLM.from_pretrained('ByteDance-Seed/Stable-DiffCoder-8B-Instruct', trust_remote_code=True, torch_dtype=torch.bfloat16).to(device).eval()
tokenizer = AutoTokenizer.from_pretrained('ByteDance-Seed/Stable-DiffCoder-8B-Instruct', trust_remote_code=True)
prompt = 'Write a quick sort algorithm.'
m = [{"role": "user", "content": prompt}, ]
prompt = tokenizer.apply_chat_template(m, add_generation_prompt=True, tokenize=False)
input_ids = tokenizer(prompt)['input_ids']
input_ids = torch.tensor(input_ids).to(device).unsqueeze(0)
out = model.generate(input_ids, steps=512, gen_length=512, block_length=4, temperature=0., remasking='low_confidence', tokenizer=tokenizer, shift=False, threshold=None, eos_id=tokenizer.eos_token_id)
print(tokenizer.decode(out[0][input_ids.shape[1]:], skip_special_tokens=True))Explore its uses
Keep exploring
Qwen2.5 Coder · 7B
Code generation, repair, and reasoning in a compact open-weight coding model.
Mistral Small 4 · 119B
Mistral Small 4 119B A6B is an Apache-2.0 licensed hybrid model from mistralai that combines general instruction, reasoning, and Devstral-style capabilities in one model. It uses a Mixture-of-Experts architecture with 119B total parameters, 6.5B activated per token, 128 experts with 4 active, and a 256k context length. It supports multimodal input (text and image) with text output, multilingual support across dozens of languages, system prompts, function calling, JSON output, and configurable reasoning effort. The model card also mentions optional efficiency resources: an eagle head for speculative decoding and an NVFP4 checkpoint. It is tagged for vLLM and offered in safetensors format.
Qwen3-Coder-Next · 80B A3B
Qwen3-Coder-Next is an open-weight text-generation model from Qwen/Qwen3-Coder-Next, licensed under apache-2.0. It is designed specifically for coding agents and local development. The model has 80B total parameters with 3B activated parameters, a 262,144-token native context length, and a hybrid architecture combining Gated DeltaNet, Gated Attention, and Mixture-of-Experts layers. It supports only non-thinking mode and does not generate thinking blocks.