rnj-1-GGUF

rnj-1 from EssentialAI is an 8.3B-parameter dense language model family trained from scratch on 8.4T tokens with a Gemma 3-like architecture featuring global attention, YaRN for 32K context extension, and 128K vocabulary, optimized for code generation, STEM tasks, math, science, and agentic workflows under Apache 2.0 license. The base rnj-1 and instruct-tuned rnj-1-instruct excel in benchmarks like HumanEval+, MBPP+, LiveCodeBench v6 (competing with larger models), SWE-bench Verified (20.8% bash-only, surpassing Gemini 2.0 Flash), MultiPL-E multilingual code, Enamel performance optimization, Berkeley Function Calling Leaderboard, GPQA-Diamond, and fill-in-the-middle (86.21% HE-FIM-Python), with strong pass@N scaling potential for specialization. Supporting vLLM/SGLang inference, llama.cpp quantization, Transformers tool-calling (Hermes parser), FIM, and integration with Cline, Claude Code, mini-SWE-agent for real-world coding assistants, PR fixes, and data analysis, it processes up to 24M token batches efficiently post mid-training/SFT phases using Muon optimizer.

rnj-1 [GGUF]

File Name	Quant Type	File Size	File Link
rnj-1.BF16.gguf	BF16	16.6 GB	Download
rnj-1.F16.gguf	F16	16.6 GB	Download
rnj-1.Q8_0.gguf	Q8_0	8.84 GB	Download

Quants Usage

(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)

Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):