🤝 Open to Collab

Steffen Röcker PRO

sroecker

·

https://x.com/sroecker

AI & ML interests

Local models

Recent Activity

liked a model about 6 hours ago

vllm-sr/d3-flash

liked a model 1 day ago

vllm-sr/d3-lite

liked a model 1 day ago

Mercity/Qwen3-8B-LaCo-30L

View all activity

Organizations

upvoted a collection 15 days ago

Typed decisions on LiteRT

Models that read a text and pick one of the options you write, each with a probability. LiteRT conversions for Android, all Apache-2.0. • 7 items • Updated 17 days ago • 1

upvoted a collection 26 days ago

Jev-style decision models: open data

Data for Jev-style decision models. Notes say where labels come from: humans, exact rules or an LLM teacher. Last updated 2026-09-24. • 17 items • Updated 26 days ago • 10

upvoted an article about 2 months ago

Article

Training a coding model to paint watercolours with TRL and OpenEnv

sergiopaniego

•

Sep 3

• 80

upvoted a paper about 2 months ago

Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090

Paper • 2608.27370 • Published Aug 27 • 40

upvoted 2 collections about 2 months ago

Puro-2B

普罗-2B • 11 items • Updated Aug 28 • 14

🌟 Antares

Agentic Models for Vulnerability Localization • 2 items • Updated Jul 21 • 62

upvoted an article 2 months ago

Article

Training a coding agent using the OpenCode harness in remote HF sandboxes with TRL and OpenEnv

sergiopaniego

•

Aug 5

• 30

upvoted 4 collections 5 months ago

🧬 Carbon

Carbon 500M, 3B, 8B genomic models and GGUF variants for llama.cpp • 7 items • Updated Jun 2 • 50

NuExtract3

12 items • Updated May 26 • 17

GSQ

GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling, https://huggingface.co/papers/2604.18556 • 10 items • Updated Aug 22 • 15

Fixed Chat Templates for Qwen 3.5 & 3.6

Rewritten Jinja templates fixing 5 bugs in official Qwen 3.5/3.6. Works in LM Studio, llama.cpp, MLX, vLLM. • 1 item • Updated Apr 30 • 13

upvoted a paper 5 months ago

Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling

Paper • 2605.13301 • Published May 13 • 166

upvoted 2 collections 5 months ago

SpecDrift

Models released as a part of Attention-Drift Paper, trained for deployment on production • 2 items • Updated May 10 • 3

Gemma 4 Assistant GGUF

Gemma 4 MTP assistant drafters as GGUF (F16/Q8_0/Q5_K_M/Q4_K_M/Q4_K_S). Speculative-decoding heads for the atomic-llama-cpp-turboquant fork. • 4 items • Updated May 7 • 15

upvoted 2 collections 6 months ago

8GB VRAM Local LLMs - Practitioner Tested

11 models on RTX 4060 Ti 8GB. speed king: Llama 1B (228 tok/s). quality king: Gemma 4 E4B (6/6). agentic king: gpt-oss-20b (10/10). • 21 items • Updated May 21 • 7

Granite 4.1 Language Models

Efficient language models for multilingual generation, coding, RAG, and AI assistant workflows. • 15 items • Updated 5 days ago • 71

upvoted an article 6 months ago

Article

Granite 4.1 LLMs: How They’re Built

ibm-granite

•

Apr 29

• 95

upvoted 3 collections 6 months ago

APEX Quants (GGUF)

MoE models quantized with the APEX Quantization technique ( https://github.com/mudler/apex-quant ) • 49 items • Updated 6 days ago • 151

1930 Coder

Fine-tuning the Talkie 13B 1930 model on agentic trajectories • 4 items • Updated May 5 • 4

Qwen3.6

4 items • Updated Apr 22 • 503