Qwen
Chinese Models Excluded by Contract: Only Four Local Document Tools Remain
A contract bars Chinese software and models, narrowing local document-processing choices and making model provenance a critical procurement issue.
Back to home
Gemma
30 articles tagged with this topic
Gemma
Gemma Loses 85% of Its Weights and Still Works: Specialized Models Are Cutting Local AI Costs
A developer pruned Gemma 4 26B-A4B to 15% of its size, retaining 92% agreement on one classification task while cutting memory use to 3.8GB.
Gemma
Google Gemma Runs on a $200 Chip — Local AI Is Leaving the Datacenter
A Reddit developer got Google's Gemma 4B running on a $200 NVIDIA Jetson Orin Nano, beating llama.cpp speeds. The hardware floor for local AI just dro
Qwen
游戏显卡就能跑的大模型排行：Qwen、GLM 进前五，但 200B 以下都在瞎编
Reddit ranks LLMs runnable on 10-16GB GPUs. Qwen and GLM crack top 5, but sub-200B models still fabricate anime, film, and music knowledge.
Gemma
16GB VRAM Drives a Real RC Car on a Local LLM — Edge AI Hits Consumer Hardware
Developer drives a real RC car with a 16GB VRAM local LLM via Ollama — edge AI finally usable on consumer hardware, complementing not replacing cloud.
LocalMind
A 37GB Model Now Fits in a Browser Tab — Even on 24GB Laptops
LocalMind by NakliTechie runs a 37GB model in-browser via WebGPU + disk-streamed MoE—even on 24GB Macs. ~3x slower than llama.cpp, MIT-licensed.
Gemma
5KB Assembly Runs a 2B-Parameter LLM — Edge AI Needs No GPU Stack
Developer tomtsai28 compressed Gemma-2B inference into 5KB of pure x86-64 assembly, hitting 4.6 tokens/sec on an old i5 desktop. Zero PyTorch, zero C
LocalLLaMA
8GB VRAM Now Runs LLMs — Qwen and Gemma Are Racing for Local AI Gamers' GPUs
Reddit's LocalLLaMA community is abuzz over running large models on 8GB VRAM—Qwen4 35B-A3B and Gemma 26B QAT lead the charge. Local AI is shifting fro
LocalLLaMA
One Workstation, 3.5 Hours, 83% Accuracy: Open-Source AI Crashes Training Costs
Reddit user firelex trained a 2B classification model in 3.5 hours on a single RTX PRO 6000 GPU, hitting 83.1% — matching commercial Jev at 83.0%.
Qwen
70 Open-Source 'Decision Small Models' Ranked — Qwen and Gemma Sweep Top 3
Qwen and Gemma dominated a benchmark of 70 open-source small models on zero-shot classification, proving local AI routing now viable without big-cloud
Google
95 Tokens for Cloud Plans, 97% Stay Local: Task Routing Outweighs the Model
Antigravity's local-model support shows task routing in action: cloud gets summaries, local runs execution. Beats bigger models on enterprise complian
Google
Google Gemma Developers Demand Updates on Reddit — Open-Source LLMs Lag Behind
Google's open-source Gemma LLM faces Reddit backlash—users demand a 220B MoE version as open vs closed-source competition intensifies.
Google
700 Lines of C Code Run Google's Latest LLM — Solo Project Beats llama.cpp
Open-source gemma4.c runs Google Gemma 4 E2B in 700 lines of C, hitting 25.9 tok/s on a regular CPU and beating llama.cpp.
Bailing
4060Ti Replaces Gemma-12B — Big Models' Side Hustle Is Being Redistributed
A Reddit user replaced Gemma-12B with Ling Tiny on a 4060Ti, calling the speed "amazing." Auxiliary AI tasks are migrating to consumer hardware.
Gemma
One Dev, 16GB GPU Triples Open-Source AI Tool Calling — Local Agents Got Cheap
This week a Reddit dev fine-tuned Google's Gemma 12B on a 16GB consumer GPU, boosting AI tool-calling 2.7×. Local Agent costs are visibly falling.
Gemma
Regular users now customize model files — local AI barrier drops a notch
Regular users now build GGUF files themselves on r/LocalLLaMA. Local AI is maturing — Chinese enterprises should rethink private deployment math.
Google
Behind Gemma's 1 Billion Downloads: Google Is Serious on Open Source — but Late
Gemma downloads hit 1 billion; Google celebrates in SF with Demis Hassabis. Yet Meta and Alibaba have already seized the open-source throne.
Qwen
Open-source LLMs: coding soars, writing stalls — are benchmarks off-track?
Qwen's coding record grabs headlines, but Reddit's open-source community warns writing and translation lag. We ask if R&D priorities have drifted.
Qwen
Qwen Overthinks, Gemma Lazy, Muse Bland — Open-Source LLMs Face the Nitpick Era
This week, Qwen 3.8, Gemma 4, and Muse Glimmer all drew Reddit user complaints. The backlash reveals real tension in the open-source LLM ecosystem.
Gemma
Small Model Coding +8.55%: Open-Source Quantization Breaks 'Bigger Is Better'
Tensor-level quantization lifts Gemma 4 12B coding scores 8.55% with negligible size gain, chipping away at the "bigger is better" LLM narrative.
Radeon 780M
A $700 Mini PC Runs 35B LLMs Locally — Budget AI Path Emerges
Reddit's LocalLLaMA community proves an AMD Radeon 780M mini PC with 64GB DDR5 can run Qwen 35B and Gemma 31B locally under €1000.
Qwen
Qwen Wins at Code, Gemma at Text — Tokenizer Gap Is the Hidden Cause
A Reddit user's test found: the same 330 lines of code became 1,609 tokens in Qwen but 4,258 in Gemma — a 2.6x gap that may explain their very differe
Gemma
Generate Three Times, Self-Select the Best: How Small Models Hit Stable Output
A developer runs Gemma 4 12B to summarize YouTube transcripts using multi-sample generation plus self-evaluation, proving small local models can deliv
Gemma
Gemma 4 Per-Layer Embeds: Knowledge-Reasoning Split, Hope or Hype
Gemma 4's per-layer embeddings spark debate: Can knowledge and reasoning scale separately? If so, 2B models could hold 20B knowledge, redefining local
hfviewer
New Hugging Face Visualizer Cracks Open AI Black Boxes Without Code
hfviewer.com visualizes Hugging Face model architectures interactively. It replaces code with intuitive graphics, lowering the barrier to grasping AI
Qwen
Qwen 3.6 Wins Benchmarks, Fails Reality: Benchmaxing Distorts AI Perception
Qwen 3.6 won benchmarks but lost to Gemma 4 in practice, burning 8000+ tokens in a loop. Benchmaxing distorts AI perception; firms must shift to real-
Gemma
Gemma 4 Hits HuggingFace — Open Source Outpaces Official Toolchain
gemma-4-31B-it-DFlash on HuggingFace lacks llama.cpp support. We see models outpacing toolchains—having models you can't run is the new paradox.
NVIDIA
NVIDIA NVFP4 Puts 26B Model on Consumer GPU With Under 1% Accuracy Loss
NVIDIA's NVFP4 Gemma-4-26B shrinks to 18.8GB for consumer GPUs with <0.7% accuracy loss. 4-bit is now optimal, but also an ecosystem lock-in.
Qwen
Gemma 4 Beats Qwen 3.6 With 1/5 The Tokens — Local AI Era Demands Efficiency
A Reddit test shows Gemma 4 beats Qwen 3.6 on a Pac-Man prompt using 1/5 the tokens and time. We argue: in local deployment, efficiency now trumps raw
Pocket LLM
手机本地跑 AI 不再需要联网—— 一个开源安卓应用正在把这件事变得可操作
Pocket LLM v 1.4.0 shrinks to ~200MB, lets users download models on demand and run AI fully offline on Android.