The Builder

One-command LLM censorship removal — now with reproducibility

“Reproducible outputs and honest benchmarking are the features that matter here — not the censorship angle. I've had local models behave differently on identical prompts due to VRAM spikes causing partial loads. Heretic 1.3 fixing that alone makes it worth running for any serious local deployment.”

1.6T open-source MoE that nearly matches frontier — MIT, 1M token context

“MIT license on a 1M context model that beats GPT-5 on coding evals is wild. V4-Flash at 13B active params is particularly practical — you get near-frontier coding performance with inference costs that don't require a mortgage. Ship immediately.”

Google's open multimodal models — vision, audio, and text under Apache 2.0

“Apache 2.0 on a model that beats GPT-class performance at 31B? Ship it immediately. The MoE 26B variant is already running under 16GB VRAM for me with llama.cpp quantization. The unified multimodal arch saves a ton of pipeline complexity.”

27B dense coding model that outperforms models 10x its size on benchmarks

“A 27B model beating a 397B model on coding benchmarks at Q4 quantization that fits on a single GPU is genuinely exciting. This changes the economics of self-hosted coding agents. I'm testing it in my agentic pipeline immediately. The Qwen team has been consistently delivering quality — this continues that trend.”

104B MoE model with only 7.4B active params — big model quality at small model speed

“7.4B active parameters at 104B capacity is the best ratio in its class right now. If the benchmark performance holds up in real workloads, this is an easy drop-in for high-throughput API use cases where cost-per-token matters. Free on OpenRouter means zero risk to test it against your current model.”

1.58-bit LLMs that run at 82 tok/s on M4 Pro and on your iPhone

“82 tokens per second on M4 Pro in 1.75 GB is a genuinely impressive engineering achievement. For local tooling, code assistants, or any latency-sensitive workload where I don't want cloud round-trips, this hits a sweet spot that larger quantized models miss. Apache 2.0 means I can embed it in commercial apps without legal headaches.”

35B total, 3B active: Alibaba's lean MoE coding beast goes fully open source

“3B active parameters with 35B parameter breadth is engineering magic. I'm getting near-frontier coding results in Cline and running it locally on a 3090 — the refusals are lower than Claude for security research too. Apache 2.0 means I can fine-tune it on my codebase. This is the best open-source coding model I've used.”

1.58-bit LLMs that fit in 1.75 GB — runs in your browser via WebGPU

“1.75 GB for an 8B model is a genuine engineering achievement. I can finally ship a capable model inside a desktop Electron app without requiring users to have a dedicated GPU. The WebGPU demo loads fast and output quality is surprisingly coherent for its size.”

First commercially licensed 1-bit LLMs — 8B in 1.15 GB, 8x faster on-device

“1.15 GB for an 8B model is the number that matters. I can run agents on a Raspberry Pi 5 now without thermal throttling. The commercial license means I can actually deploy this in products — that was always the missing piece with research-only 1-bit work.”

3B-parameter open model supporting 70+ languages — runs offline on a phone

“Ollama support means this is running locally in ten minutes. The region-specific variants are a smart design choice — a model tuned for South Asian languages will outperform a globally averaged model on those languages even at smaller parameter counts. This is the right architecture for the problem.”

1-bit quantized 8B LLM — 1.15GB, runs on-device at 368 tok/s

“1.15GB for an 8B model that runs at 368 tok/s is genuinely remarkable. Fitting LLM intelligence into a package that runs on a phone CPU opens use cases that were completely impractical months ago. For offline apps, robotics, or privacy-sensitive deployments, this changes the calculus entirely.”

399B open MoE reasoning model that's 96% cheaper than Claude Opus

“Near-Opus-level reasoning at $0.90/M tokens is the pricing inflection I've been waiting for. Apache 2.0 weights mean I can self-host for compliance-sensitive use cases. Already benchmarking it as a drop-in for my agent evaluation pipeline.”

Google's first Apache 2.0 open model family with native multimodal

“Apache 2.0 means I can embed it in commercial products without legal review overhead. Native audio + 256K context on a 26B model that runs on a single A100 is a killer combo for production agent work. This is the open model I've been waiting for.”