Open-Weights

Matt Wolfe edits an animated image, tests hosted video demos, and finds the practical limits behind speed claims, open weights, and free allowances.

IBM’s Mixture of Experts separates targeted model repair from recursive self-improvement, then examines open weights and the limits of world-model simulations.

Cole Medin tests three coding-factory configurations, finds costly rework in long runs, and keeps open models on scoped implementation.

A local model identifies three small networks’ training tasks from their numerical weights. The failed fourth case and hour-long runs keep the claim grounded.

MIT researchers report targeted weight edits that let a small Qwen model finish more answers, plus experiments reading and writing game networks.

Matt Wolfe reviews Anthropic's proposed IPO, model releases, a hands-on Genie 4 demo, and an AI-generated complexity-theory claim.

IBM’s Grace Ableidinger explains lower-precision model storage, the memory arithmetic, and the tradeoffs between post-training quantization and quantization-aware training.

David Abitbol reports Revolut’s experience serving 2,000 engineers with GLM models: lower token bills, faster delivery, and substantial GPU-serving work.

Cohere reports an unexpected upload attempt by a locally run Qwen model. The sandbox blocked it; the cause remains unresolved.

MacPaw researchers test four families of model-serving optimizations on Apple Silicon. Their measurements show why a server-GPU speedup needs checking against the model, software and request pattern you actually use. Pruning the KV cache, a …

Directable voices, persistent AI worlds and 3D reconstruction, with the hardware limits and demo-only claims kept in view.

A practical case for separating the model you use from the company you trust—with HN pushback on privacy, cost, and coding quality.

A month of cheap coding-agent work suggests that 'good enough' changes what you delegate; practitioners disagree on where it stops being good enough.

A 2-billion-token experiment found that cheap open models can handle routine coding—but model mix, provider capacity, and unbudgeted prototypes dominate the bill.

Open Chinese model research may be cutting the cost of long-context AI — but lower prices are not proof that Western labs copied the underlying design.

Medin hit the $200 Claude Max and $200 Codex Pro limits days before reset; the fix is model routing — most capable model for planning and review, cheap open model for implementation, roughly 4x cheaper with the same or better output.

A CMU professor argues that cheap-to-run local models and autonomous agents are about to hand the next decade of AI research to university labs — and the 94-comment thread is mostly about whether the claims add up.

A GPU-rental data firm argues that the standard "new generation kills the old one" depreciation model is wrong — because open weights give older silicon a workload it can still win.

Open model weights alone do not give countries or institutions control over their AI. The dependencies run through chips, cloud, agent harnesses, data systems, and observability.

An essay reads this summer's AI panic as a sales pitch: the labs want an antitrust waiver and a regulatory moat, because their real fear is free open-weight models.

A year-old open model, a funded lab's 'breakthrough', and the measurement that matters: a classifier's own confidence never warns you when it cannot read the input.

A $1,200 experiment in learning Postgres query hints, with HN's objections about planning costs, tuned baselines, and what transfers to production.

Anthropic's CEO asks for open-weight regulation, a distillation crackdown and an antitrust waiver. A security engineer reads the essay closely and checks the one threat claim that sits in his own field.

Armin Ronacher accepts Dario Amodei's account of the danger and rejects the fix: there is nothing to pace, and open weights are the pacing mechanism. Plus what 114 HN commenters added.

A reply to Anthropic's call to pace the frontier: every regulation the labs ask for ends in capture, so Jake Gold proposes a rule with nothing to game — any model offered to the public ships as open weights.

The New York Times reports that companies are moving routine AI work to cheaper open models, while reserving expensive closed systems for demanding tasks.

One developer's case for owning your inference: a 48GB Mac mini running open-weight models beats rented cloud APIs on cost, privacy, and control.

Your local model isn't dumber — the inference stack quietly changes its outputs: attention backends, KV-cache quantization, and weight quants flip tokens and break tool calls.

A 17GB open-weights model that rivals much larger systems — if you dial down its comically aggressive default reasoning effort.

Z.ai shows that scaling RL post-training on diverse, long-horizon coding environments produces rapid gains — and an unexpected side effect: emergent cyber exploitation capability.

Zach Mueller (Lambda) on when open models beat APIs: the current open-model landscape, token-speed targets, quantization trade-offs, and why open weights are good enough for ~90% of tasks for ~90% of people.

Mistral releases Shieldstral: a 3B open-weights multimodal safety classifier that takes your moderation policy as a plain-language question at inference time — Apache 2.0, runs on a 16GB GPU.

ComfyUI ships day-zero support for MiniMax H3: an omni-modal video model with native stereo audio, 2K output — optimized to run locally on an RTX 3060.

Wolfe tests Locally AI with Qwen 3.5 on an iPhone, showing an airplane-mode response alongside reasoning failures, heat and slower long chats.