Opening

This week's radar picked up three "download-and-run" highlights: KRAFTON released a 3B bilingual model with 262K context, ConceptCountry halved 27B inference tokens via weight-space editing while getting stronger, and MinimaLabs dropped a 99M parameter model that's the new go-to for fine-tuning tutorials. I'll tell you which one fits tonight's hack session.

This Week's Picks

1. KRAFTON/Raon-1-nano: 3B Params, 262K Context, Native Mac Bilingual Tool Calling

• What it is: A 3B dense model distilled and pruned from KRAFTON's in-house 30B-A4B MoE. The official model card claims 262K context, supports mixed Traditional Chinese/English chat and tool calling, licensed Apache-2.0.
• Where it fits:
  • Running a local bilingual assistant with tool calling on your MacBook (Apple Silicon) without sending data to the cloud.
  • Corporate security policies forbid external APIs, but you need long-document summarization or code assistance, 3B + 262K context handles it on a single machine.
  • Deploying a voice assistant backend on a laptop or mini PC: run inference with GGUF via llama.cpp, no extra GPU needed.
• Can it run?: The model card lists BF16 (~6.3 GB, unquantized), GGUF (llama.cpp), MLX 4-bit GPTQ (Apple Silicon), and LiteRT-LM (4-bit weights, fp16 activations, 4K tokens). Mac users can load the MLX 4-bit quantized version directly for inference.
• Buzz this week: HF trending score 92 (1.3k downloads, 94 likes).
• Versus peers: The model card claims it leads same-size models across 17 benchmarks (e.g., Korean chat 91.0). Compared to its predecessor Raon-1 mini preview (30B-A4B MoE), pruning and distillation drastically shrank the model while retaining tool calling and bilingual abilities.
• How to get it: README provides full example code for Transformers, vLLM, llama.cpp, MLX, and LiteRT-LM. Mac users can load the MLX 4-bit quantized version directly.
• Link: https://huggingface.co/KRAFTON/Raon-1-nano

2. ConceptCountry/qwen-concept-27b: Weight-Space Editing Qwen2.5-27B, No Fine-Tuning, No Prompt Engineering, 27-54% Fewer Reasoning Tokens

• What it is: ConceptCountry manually edited nine down-projection matrices in Qwen2.5-27B's weight space (the model card notes it's based on Qwen2.5-27B; a source reference to Qwen3.8-27B is a typo). No fine-tuning, no repacking, no prompt engineering. They published the paper, data, and fully reproducible code for every number. Licensed Apache-2.0.
• Where it fits:
  • Researchers or advanced practitioners wanting to understand whether "edit weights directly, no training" can improve model behavior, download the code and reproduce every figure in the paper.
  • Running Qwen2.5-27B lighter with more efficient reasoning (27-54% fewer thinking tokens). The Q4_K_M quant (16 GB) runs on a single 24 GB consumer GPU.
  • Tech reading groups or teaching "weight-space programming", the middle ground between fine-tuning and prompt engineering, using this as a living case study.
• Can it run?: Official formats: BF16 (52 GB), FP8 (30 GB), Q8_0 (27 GB), Q4_K_M (16 GB). The 16 GB quant fits on a single consumer GPU.
• Buzz this week: HF trending score 198 (64.0k downloads, 263 likes).
• Versus peers: The authors ran a "norm-matched random edit" control, proving gains come from precise editing, not random perturbation: +4.3 on GPQA Diamond, +13.0 on LiveCodeBench v6. They flag this as still experimental, under evaluation.
• How to get it: The model card notes support for llama.cpp, Ollama, LM Studio, and links to the paper, data, and "code that generates every number" (concept.country/research/weight-space-programming).
• Link: https://huggingface.co/ConceptCountry/qwen-concept-27b

3. MinimaLabs/KLM-1.6: 99M Param Gated DeltaNet-2 Base Model, Runs on CPU, Fine-Tunes in Hours

• What it is: MinimaLabs' 99M parameter tiny base model using Gated DeltaNet-2 hybrid attention. Official stats: #1 Intelligence Index (33.95) on the Open SLM leaderboard for sub-100M models. Licensed Apache-2.0. This is a base model, no instruction tuning applied.
• Where it fits:
  • Running a base LLM entirely on CPU on a Raspberry Pi or old laptop, fine-tuning on your own domain data for lightweight experiments.
  • Teaching "fine-tune an LLM from scratch", 99M params is small enough to finish fine-tuning in hours on a consumer GPU or even CPU.
  • Studying how new attention architectures like Gated DeltaNet-2 behave on small models, use this model's config as a reference.
• Can it run?: Official note: "Runs on CPU and any GPU, no extra packages beyond Transformers." 99M params, tiny model files, consumer hardware handles it easily.
• Buzz this week: HF trending score 81 (1.7k downloads, 81 likes).
• Versus peers: Official claim: #1 Intelligence Index (33.95) among sub-100M models on Open SLM. But it's an untuned base model, needs instruction tuning to chat like an assistant. Compared to already-chat-tuned small models (like KRAFTON's Raon series), KLM-1.6 is better suited as a research and fine-tuning starting point, not a drop-in assistant.
• How to get it: Official Python examples cover basic usage and batched inference (fast_generate). Load directly with Transformers, no extra dependencies.
• Link: https://huggingface.co/MinimaLabs/KLM-1.6

Closing

These three models map perfectly to three on-device entry points: Raon-1-nano solves the Mac user's need for bilingual, long-context, tool-calling without leaving the laptop; qwen-concept-27b uses "weight surgery" to leap inference efficiency, giving GPU-budget engineers a no-retrain shortcut; KLM-1.6 drops the fine-tuning barrier to near zero, ideal as a teaching or architecture-experiment launchpad. If you want results tonight and have a Mac, grab Raon-1-nano's MLX build, fastest to first token. See you next week.