730B Parameters Still Do Not Make GLM-5.3 Local
GLM-5.3 is open weight, but 141 shards and a million-token ceiling make local deployment an infrastructure decision, not a laptop install.
TOPIC_INDEX
models
58 published entries in this topic.
MODELS_ENTRIES
RSS← ALL_TOPICSA 4.5 Cent Agent Model Still Needs Serious Hardware
GLM-5.3-Flash makes hosted agent work look cheap at $0.045 per task, but its 320B total parameters change the local deployment math.
16GB Is Enough for Qwen3.8 If You Pick the Right Quant
Unsloth Dynamic 3.0 makes 16GB Qwen3.8-27B deployments plausible, but only if you choose the quant for memory headroom and test real coding tasks.
The Open-Weight Wait Changes GLM-5.3's Value
GLM-5.3 looks impressive, but its missing weights and unsettled API change the decision. Here is who should test it now and who should wait.
The 3D world generator is a pipeline, not a model
WorldClaw turns one prompt into editable 3D scenes, but the public evidence still describes a research pipeline rather than a production-ready dependency.
Open weights do not make Ling 3.0 Flash small
Ling-3.0-flash is open weight, fast, and still a 124B model. Here is what 5.1B active parameters really change for self-hosted agents.
A Quarter the Size Still Hits 80% on SWE-bench
Inkling-Small uses 12B active parameters to approach its 975B sibling on coding and reasoning, but its factuality gap still matters.
82.7 Makes Cheap AI Coding Hard to Ignore
DeepSeek V4 Flash 0731 posts an 82.7 on Terminal Bench 2.1 and brings cheap, Codex-ready agent coding to public beta.
Four Billion Parameters Beat the Big Image Models on Speed
Microsoft's 4B Mage-Flow pairs a cheap tokenizer with native-resolution diffusion, hitting 0.59 seconds at 1024² on one A100 while exposing the tradeoffs.
Half the price is easy. Token efficiency is the story.
Claude Opus 5 keeps Opus 4.8 pricing, adds adaptive effort controls, and targets cheaper long-running agent work.
Open weights just reached 2.8 trillion parameters
Moonshot AI's 2.8T-parameter Kimi K3 is close to the closed frontier, but the real test starts when its promised weights arrive.
Video generation now has to listen
Vidu S1 targets live voice-controlled video at 25 FPS, with a reported 42 FPS peak on an RTX 5090.
Three Gemini models shipped, but Pro is still missing
Google shipped three Gemini models but the promised Pro update remains missing as Gemini 4 pre-training begins. 3.6 Flash delivers 17% fewer tokens and better coding scores.
2.4 trillion parameters and no benchmarks to prove it
Alibaba claims Qwen 3.8 is second only to Claude Fable 5. But there are no benchmarks, no model card, and no open weights yet. Here is what we actually know.
2.8 trillion open-source parameters just changed the AI race
Moonshot AI dropped Kimi K3, a 2.8-trillion-parameter open model that matches Claude and GPT on key agentic benchmarks. Open weights arrive July 27.
That 27B model was too big for a phone. Not anymore.
PrismML's Bonsai 27B compresses a full 27B-class reasoning model to 3.9 GB, running on an iPhone 17 Pro at 11 tok/s. The tradeoffs are real, but so is the trajectory.
$4.40 per million tokens just matched the $200 tier
A 744B-parameter open-weight model from Beijing lands within a point of Claude Opus 4.8 on agentic benchmarks at one-fifth the cost. The margin mirage is over.
Production AI build times just got cut in half
Ploy published every number from moving their production AI agent from Claude Opus to GPT-5.6 Sol. The result: 2.2x faster builds, 27% lower cost, better visual scores, and three hard-won lessons.
A coding model trained on 2 trillion interactions just undercut everyone
xAI's Grok 4.5 was trained on 2 trillion tokens of Cursor developer interactions. It costs $2 per million input tokens, uses 4.2x fewer tokens than Opus 4.8, and trades benchmark leadership for raw cost-per-task efficiency.
That $1.25 API price just forced Anthropic and OpenAI to respond
Meta's Muse Spark 1.1 API costs $1.25 per million input tokens, undercutting Claude and GPT by 75-80%. The real story isn't the benchmarks.
The AI voice interrupt problem just got fixed
OpenAI shipped GPT-Live, a full-duplex voice model that listens and speaks simultaneously. The awkward pause is finally dead.
Only 20 companies got access to GPT-5.6. The government decided.
OpenAI's GPT-5.6 Sol benchmarks competitively with Mythos but only 20 government-approved companies can use it. Three tiers, Cerebras at 750 TPS, and a regulatory framework that changes everything.
That $2/M intro price is hiding a 30% tokenizer tax
Claude Sonnet 5's $2/M intro price hides a 30% tokenizer inflation that makes it cost more per task than Opus 4.8.
Stop overthinking: a coding model just cut reasoning tokens by 30%
Moonshot AI says K2.7 Code cuts reasoning tokens by 30%. Independent testers aren't convinced by the benchmarks. The pricing math still works.
Meituan's open model matched GPT-5.5. They deliver food for a living.
Meituan's 1.6T open model matched GPT-5.5 on SWE-bench Pro, trained on 50K domestic cards. MIT license, $0.75/M input tokens.
A 753B open-source model just matched Claude on coding
Zhipu AI's GLM-5.2 is a 753B open-weight MoE model under MIT license that matches Claude Opus on coding benchmarks, with a genuine 1M token context window and $18/month flat-rate pricing.
Your coding agent's harness is the bottleneck it can't see
DeepReinforce's Ornith-1.0 learns to write its own RL scaffolds, and the 35B model beats Qwen 3.5-397B on terminal tasks.
91.9% coding score. The government decides who uses it.
OpenAI shipped GPT-5.6 Sol with a 91.9% coding score. Almost nobody can use it. The US government controls who gets access.
That expensive API subscription just lost to a free download
GLM-5.2, a 753B open-weights model from Z.ai, just beat GPT-5.5 on multiple coding benchmarks at one-sixth the cost under an MIT license.
That 72-Hour AI Model Got Banned by the Government
Claude Fable 5 lasted exactly three days before the US government forced Anthropic to pull it. The first model killed by export controls sets a permanent precedent.
Your iPhone can't run Apple's new 20B AI model
Apple announced AFM 3 at WWDC26: a 20B sparse on-device model using Instruction-Following Pruning. Only iPhone 17 Pro/Max/Air can run it. The cloud tier depends on Google's NVIDIA GPUs and Gemini distillation.
One token at a time is a bottleneck. Google broke it.
Google released DiffusionGemma, a 26B open model that generates text 4x faster by replacing token-by-token decoding with parallel diffusion. The speed is real. The quality gap is too.
Zhipu just shipped 1M context to coding. Where are the benchmarks?
Zhipu shipped GLM 5.2 with 1M token context to their coding plan today. No benchmarks. No independent verification. MIT weights coming next week. Here's what we know and what we don't.
Your coding agent wastes tokens thinking. This one doesn't.
Moonshot AI released Kimi K2.7-Code, a 1T-parameter open-source coding model that cuts reasoning tokens by 30% at $0.95/M input. The real story is the economics, not the benchmarks.
Anthropic's strongest model has two versions. You get the weaker one.
Anthropic released Claude Fable 5 and Mythos 5: the same model, split by safety tier. The benchmarks you see everywhere mostly belong to the version you cannot buy.
Open-weight image models finally stopped losing to closed ones
Ideogram 4.0 is the first open-weight image model to top the DesignArena leaderboard, with 0.97 OCR accuracy and structured JSON prompting that gives designers precise layout control. It runs on a single 24GB GPU.
US Open-Weight AI Finally Has a Fast One
NVIDIA's Nemotron 3 Ultra is the fastest US open-weight model at 300+ tokens per second, built specifically for long-running agents. 550B parameters, 55B active, hybrid Mamba-Transformer architecture.
NVIDIA collapsed 5 AI models into one. Robots just got cheaper.
NVIDIA released Cosmos 3, an open omnimodel that merges vision reasoning, world generation, and robot action prediction into a single architecture. The five-model stack for physical AI may be obsolete.
Your 3070 runs a 35B model at 22 tokens per second
MoE architecture and llama.cpp turned a 2020 gaming card into a local AI workstation. Here's the real hardware data.
Your laptop runs multimodal AI now. Nobody told you.
Google DeepMind released Gemma 4 12B, an encoder-free multimodal model that runs on 16GB laptops and processes text, images, audio, and video through a single transformer. Apache 2.0.
Your phone can now run FLUX image generation locally
PrismML's Bonsai Image 4B compresses FLUX.2 Klein down to 0.93GB using 1-bit quantization, enabling full local image generation on iPhones, MacBooks, and in-browser via WebGPU. Apache 2.0.
Your laptop just got 128K context. At 253 tokens per second.
Liquid AI's LFM2.5-8B-A1B brings 128K context to consumer hardware — 1.5B active parameters, 253 tok/s on a laptop, and tool calling that finally feels interactive.
On-device AI agents no longer cap out at 32K context
Liquid AI's LFM2.5-8B-A1B packs 128K context and real tool calling into 1.5B active parameters. It runs at 253 tok/s on a laptop. Local agents just became practical.
Nobody was watching this model until it crushed OpenRouter
Tencent's Hy3 preview has quietly become the most-used model on OpenRouter. Here's the tech, the pricing, and what it says about AI in 2026.
Chinese open-weight models are eating Silicon Valley's lunch
Chinese open-weight models now handle over 60% of tokens on OpenRouter, up from 2% a year ago. Xiaomi, Alibaba, DeepSeek, and others are winning on price-performance at 10-20x lower cost for comparable quality. Here is how it happened and what it means for developers.
Enterprise-grade AI that actually runs on premises, not just in theory
Cohere released Command A+, a 218B MoE model under Apache 2.0 that runs on two H100s. It's their bid to make sovereign enterprise AI practical , frontier-grade model, no vendor lock-in, deployable on premises or air-gapped.
Google Abandoned the Bigger Model Race for Speed
Google I/O 2026 shifted focus from bigger models to faster ones, launching Gemini 3.5 Flash, Omni for video generation, and Antigravity 2.0 for always-on AI agents.
Your $2,600 RULER Run Just Cost $8
A Miami startup just claimed 1,000x efficiency gain with a new attention architecture. 13 employees, 11 PhDs, $29M seed, and zero public access. I want it to be real. I also remember Magic.dev.
Google Hid a 3x Speedup in Gemma 4. The Community Found It in Three Days.
A Reddit user found Multi-Token Prediction heads hidden inside Gemma 4's model weights. Google said they were 'removed on purpose.' Then Google officially released them anyway. Here's how MTP works and why it matters.
Your uncensoring technique matters more than the model you modify
Gemma 4 31B and Qwen3.6 27B both had their refusal mechanisms removed — but with completely different abliteration methods. Forensic analysis shows technique matters more than model.
Grok 4.3: xAI's Heavy Multi-Agent Engine
xAI's Grok 4.3 beta runs a 16-agent architecture at 209 tok/s with up to 2M token context—competitive intelligence at a quarter of Claude's input cost.
IBM Granite 4.1: The 8B Dense Model That's Outperforming 32B MoEs
IBM's Granite 4.1 8B dense model matches 32B MoE benchmarks while running on consumer hardware. 20x token efficiency vs Qwen.
Gemma 4: Google's Open Multimodal That Runs on Your Phone
Google DeepMind's Gemma 4 brings frontier multimodal AI to phones. Apache 2.0 licensed, 140+ languages, thinking mode included.
WebGen-R1: 7B Model Rivals DeepSeek-R1-671B for Website Generation
A 7B model trained with RL achieves DeepSeek-R1-671B level website generation using scaffold-driven generation and cascaded multimodal rewards.
GPT-5.5: OpenAI's Smartest Model for Real Work
GPT-5.5 released April 23, 2026. 82.7% Terminal-Bench, 84.9% GDPval—beats GPT-5.4, Claude Opus 4.7, Gemini 3.1 Pro. Matches GPT-5.4 latency while delivering higher intelligence. Fewer tokens for same tasks.
DeepSeek V4: Million-Token Context Intelligence
DeepSeek V4 just dropped: 1.6T MoE with 1M context, MIT license. Beats GPT-5 and Gemini on LiveCodeBench (93.5%). 27% FLOPs efficiency vs V3.2. Open weights, aggressive pricing.
Qwen3.6-35B-A3B: 9x Efficiency With 35B MoE
Alibaba's Qwen3.6-35B-A3B delivers Qwen3.5-27B performance with 3B active parameters. SWE-bench 73.4%, Apache 2.0, runs at 100 t/s on M5 Max.
LLaDA2.0-Uni: First Diffusion LLM to Close Gap with Specialists
Inclusion AI's 16B MoE diffusion LLM unifies multimodal understanding and generation, closing the gap with specialized models for the first time.