Tagged with

22 articles found

Star Elastic: NVIDIA’s Model Compression Gambit That Actually Makes Sense

NVIDIA packs 30B, 23B, and 12B reasoning models into one checkpoint, achieving 360x training cost reduction and dynamic speed scaling.

#moe#nvidia#reasoning-llms...

dense

The Parameter War is Over: Why Your LLM Size Fetish is Pointless

Analysis of Qwen3.6-27B vs Coder-Next shows statistical ties despite massive parameter differences. The era of bigger-is-better has ended.

#dense#moe

agentic coding

3 Billion Active Parameters Just Challenged 30 Billion: Inside Qwen3.6’s Sparse MoE

Alibaba’s Qwen3.6-35B-A3B activates only 3B parameters per token yet claims agentic coding parity with models 10x its size. We dissect the architecture, benchmarks, and whether this Apache 2.0 release actually changes the local AI equation.

#agentic coding#alibaba#moe...

moe

Nvidia Nemotron 3 Super: The 4-Bit Trojan Horse Disguised as an Open Model

Why Nvidia’s hybrid Mamba-MoE architecture is less about benchmark glory and more about owning the inference stack from cloud to workstation.

#moe#nemotron#nvidia

blackwell

Blackwell’s 99KB Cage: How One Developer Jailbroke Qwen3.5 Performance with a 64-Line Kernel Patch

Technical deep dive into unlocking 2x inference speed on RTX PRO 6000 Blackwell GPUs by fixing CUTLASS SMEM overflow bugs for MoE models

#blackwell#cuda#CUTLASS...

moe

The 4B Model That Eats GPT-4’s Lunch: How Qwen 3.5 Rewrote the Edge AI Playbook

Qwen 3.5’s sub-10B models are outperforming last generation’s giants, and with Unsloth’s Dynamic 2.0 quantization, they’re running on your phone at 60 tokens per second. The ‘GPU poor’ just got their revenge.

#moe#quantization#qwen...

lfm2

Liquid AI’s LFM2-24B: A 24B-Parameter MoE That Runs on 32GB RAM and Makes Cloud APIs Look Overpriced

Liquid AI’s new sparse MoE model activates only 2.3B parameters per token, delivering server-class AI performance on consumer hardware while challenging the cloud-only paradigm.

#lfm2#liquid-ai#moe

minimax

MiniMax-2.5: The 230B Open Model Running on 101GB That Makes Claude Opus Look Overpriced

MiniMax-2.5 achieves 80.2% on SWE-Bench Verified with 200K context, runs locally at 3-bit precision, and costs $1/hour, forcing a reckoning for proprietary AI pricing.

#minimax#moe#open-source-ai...

ai-models

MiniMax M2.5: The $1/Hour Model That Makes Claude Opus Look Overpriced

MiniMax’s 230B MoE model hits 80.2% on SWE-Bench at 1/20th the cost of competitors. Here’s why the AI pricing model just collapsed.

#ai-models#coding-agents#m2.5...

consumer-gpu

Unsloth’s MoE Coup: The 12x Speedup That Kills the VRAM Arms Race

Unsloth’s custom Triton kernels deliver 12x faster MoE training with 35% less VRAM, enabling Qwen3 and DeepSeek fine-tuning on consumer GPUs. But the real story is what this means for AI democratization and hardware vendor lock-in.

#consumer-gpu#deepseek#Fine-tuning...

AI4Science

A Trillion Parameters and a Single Purpose: How Intern-S1-Pro Reshapes Scientific Reasoning

Intern-S1-Pro’s 1T MoE architecture delivers SOTA scientific reasoning while activating only 22B parameters, challenging closed-source models and redefining specialized AI for chemistry, materials, and life sciences.

#AI4Science#moe#multimodal...

efficiency

Step-3.5-Flash: The 196B Parameter Model That Makes Giants Look Wasteful

Stepfun’s sparse MoE model activates only 11B parameters yet outperforms models 3-5x larger on coding and agentic tasks, delivering 100-300 tok/s on consumer hardware and forcing a reckoning with the parameter count arms race.

#efficiency#moe#sparse-activation...

moe

MOVA Breaks the Silent Era of Open-Source Video Generation, And It’s Not Asking Permission

OpenMOSS’s MOVA delivers synchronized video-audio generation with 18B active parameters, challenging closed models with fully open weights and day-0 SGLang support.

#moe#multimodal#sglang...

huggingface

Hugging Face’s Transformers v5 Delivers 11x MoE Speedups by Admitting They Were Doing It Wrong All Along

Transformers v5’s 6x-11x performance gains for Mixture-of-Experts models reveal more about v4’s limitations than v5’s innovations. The API simplification and dynamic weight loading rewrite the rules for LLM inference.

#huggingface#moe

glm-47

GLM-4.7-Flash: The Reasoning Model That Can’t Stop Thinking

Z.ai’s new 30B MoE model promises transparent step-by-step reasoning, but its meticulous thought process reveals a deeper tension in local AI deployment: when interpretability becomes a performance bottleneck.

#glm-47#moe#reasoning-models

agentic workflows

GLM-4.7-Flash: The Local LLM That Actually Does What It Promises (Mostly)

GLM-4.7-Flash is delivering reliable agentic performance on consumer hardware, but the path to getting it running reveals the messy reality of local AI deployment.

#agentic workflows#GLM-4.7#llama.cpp...

mamba

Nemotron-3-nano 30B Outperforms Llama 3.3 70B: The Local LLM Efficiency Breakdown

A 30-billion-parameter model is beating Llama 3.3 70B on reasoning tasks while using a fraction of the compute. Here’s how NVIDIA’s hybrid architecture changes the local AI game.

#mamba#moe#nemotron...

benchmarks

Xiaomi’s MiMo-V2-Flash: The 309B-Parameter Underdog Giving GPT-5 a Run for Its Money

An in-depth look at how Xiaomi’s modestly-sized MoE model delivers elite performance at a fraction of the cost, and why the community isn’t buying it.

#benchmarks#LLM#moe...

local-llms

NVIDIA’s Nemotron-3-Nano: A 30B Hybrid Reasoning Model That Actually Delivers 1M Context (Mostly)

NVIDIA’s new open-weight Nemotron-3-Nano promises 1M token context and best-in-class reasoning performance, but early deployments reveal a more complicated reality. Here’s what the benchmarks don’t tell you.

#local-llms#moe#nemotron...

diffusion

Diffusion Language Models Break the Autoregressive Cage – And LLaDA2.0 is Jangling the Keys

LLaDA2.0’s MoE-powered diffusion architecture challenges everything we know about local AI deployment

#diffusion#llama.cpp#local-ai...

cerebras

When Less Is Actually More: Cerebras’ REAP Exposes Expert Merging as Flawed MoE Strategy

REAP pruning outperforms merging in MoE models, enabling near-lossless compression of 480B giants to local hardware

#cerebras#compression#LLM...

LLM

Qwen Next Just Made Every Other Local LLM Look Obsolete

Alibaba’s hybrid MoE architecture delivers 80B parameter performance with 3B activation costs, revolutionizing local task automation

#LLM#local-llm#moe