开发者对 OpenBMB MiniCPM5-1B 进行微调，发布 657MB 本地思考模型

发生了什么

Technology AI Shorts Artificial Intelligence Applications Editors Pick Language Model Large Language Model Machine Learning New Releases Software Engineering Staff Tech News A community developer, GnLOLot , has published a 1B model that runs fully on local hardware. The model is MiniCPM5-1B-Claude-Opus-Fable5-Thinking , with GGUF builds for llama.cpp-compatible runtimes. It needs no API key and makes no cloud calls. The Proposed Model The model is built on openbmb/MiniCPM5-1B . That base is a real, documented release from OpenBMB . It is a dense 1.08B-parameter model using a standard LlamaForCausalLM architecture. It has 24 layers, grouped-query attention, and a 131,072-token context length. OpenBMB reports 1B-class open-source SOTA within its own comparison set. The base already ships a native thinking template. Reasoning is toggled through enable_thinking , giving both a Think and a No Think mode. The derivative model keeps that template and MiniCPM5’s tool-call format. On top of that base, the developer applied a fine-tune. The card states the model is ‘further fine-tuned on Fable 5 data’ to improve coding and instruction following. The GGUF card repeats this as ‘post-trained on Fable 5 data.’ How it is actually built The described method is not classical distillation. You do not shrink the original model. Instead you generate many conversations with a teacher model. You capture its replies and reasoning traces as text. You then supervised-fine-tune a smaller base model on those traces. This distinction is important for accuracy. Classical distillation transfers signal from a teacher’s logits or weights. No one has access to Claude’s weights or logits. So this is supervised fine-tuning on generated outputs, not weight-level distillation. OpenBMB’s own base model, by contrast, uses a documented On-Policy Distillation stage between its own teacher and student checkpoints. The practical effect is that the 1B model learns to imitate response format and style. It does not absorb the teacher’s underlying capability. A 1B parameter budget cannot hold frontier-scale reasoning. The specs that check out The context window is 128K tokens, inherited from the base config.json (131,072). The GGUF repository ships four quantizations. Q4_K_M is roughly 657MB and is labeled the smallest footprint. Q5_K_M is roughly 751MB. Q8_0 is roughly 1.1GB and is the maintainer’s recommended default. F16 is roughly 2.1GB. The ‘657MB footprint’ is the smallest quant, not the default build. The model loads directly in llama.cpp, Ollama , LM Studio, jan, and KoboldCpp. Interactive: how the build works The explainer below walks the build pipeline, the footprint tradeoffs, and the honest split of what a fine-tune can and cannot carry over. How to run it The GGUF card gives a one-line path through Ollama: Copy Code Copied Use a different Browser ollama run hf.co/GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-Thinking-GGUF:Q4_K_M The same repository documents llama.cpp, LM Studio, jan, and KoboldCpp. Recommended sampling for Think mode is temperature=0.9, top_p=0.95 . The model may emit reasoning blocks before the final answer, which downstream apps can strip. Key Takeaways The model is a supervised fine-tune of OpenBMB’s MiniCPM5-1B on Claude Fable 5 traces, not a weight-level distillation. Real specs: 128K context, GGUF quants from ~657MB (Q4_K_M) to ~2.1GB (F16), Q8_0 the recommended default. Fine-tuning on outputs transfers format and style, not frontier reasoning or broad knowledge. No benchmarks or training dataset are published, so capability claims are currently unverifiable. Apache-2.0 covers the base weights only; training on Claude outputs raises a licensing question the card leaves open. Michal Sutter + posts Bio Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights. Michal Sutter Best Local LLMs You Can Run on a Single 24GB GPU in 2026: Qwen, Gemma, Mistral, DeepSeek Compared Michal Sutter 10 Open-Source No-Code AI Platforms for Building LLM Apps, RAG Systems, and AI Agents Michal Sutter Kimi K3 vs DeepSeek V4 Pro vs GLM-5.2: Open Trillion-Scale MoE Models Compared on Benchmarks, License, and Serving Cost Michal Sutter Google Cloud’s Always-On Memory Agent Replaces RAG and Embeddings With Continuous LLM Consolidation on Gemini 3.1 Flash-Lite

为什么重要

Google 相关变化会影响地区可用性、合规成本和供应商风险。

影响谁

Claude · Gemini · Qwen · DeepSeek

建议动作

Upgrade or review the changed version

地区可用性

No region-specific availability verified.

Generated from event evidence.

• src-marktechpost · summary
• src-marktechpost · reason