AI News
- AI News5 min
Open Weight Models: Mistral Large 4 Has No Revenue Cap, but Its License Turns on Undefined Words
By Petru Popa ·Mistral released the weights of Mistral Large 4, a mixture-of-experts model of about one trillion parameters with a 1M-token context. The license has no revenue threshold, unlike Mistral's Timestral license, but it requires a separate commercial agreement for anyone giving third parties meaningful control over inputs or parameters, and it never defines that phrase. The hardware floor is an eight-GPU node, or a 512 GB Apple machine at two bits.
Read - AI News6 min
Claude Code Token Cost: a Fifth of Requests Write a Three-Word Status Line
By Petru Popa ·A gist published on 26 October claims Claude Code spends 15 to 20 percent of tokens on a request that only updates each subagent's status line. Three earlier measurements in the Claude Code issue tracker agree. We read the 2.1.312 binary: the fork is on by default since 2.1.232, refreshes every 30 seconds at the clamp minimum, has a cheaper short mode nobody can select, and is listed on no docs page. Here is how to measure it and what the two documented levers cost you.
Read - AI News6 min
The Cheapest LLM API for Agents Is Decided by Cache Writes, Not the Price Cuts
By Petru Popa ·Four months of cuts left Claude Sonnet 5.5 and GPT-6.1 Sol at the same $2 input, $10 output and $0.10 cache-read price. The cuts landed mostly on cache reads. On a model that bills reads at 5% of input, a cache write costs 25 times a read, so in Cloudflare's own reviewer about 6% of cached tokens become 62% of the cached-context bill. Measure your write share before you switch vendors on list price.
Read - AI News6 min
Self-Hosting Beam 501B: The Weights Are Open. The Release Channels Are Not There Yet.
By Petru Popa ·Beam is a 501B mixture-of-experts model with 23B active parameters, an Apache 2.0 license and a 1M-token context. The launch posts say you can run it on your own hardware from day one. The artifacts say something narrower: transformers 5.20.1 is the only released package that knows the architecture, vLLM's support is a branch-built Docker tag over two open pull requests, llama.cpp merged support nine hours after the GGUFs were uploaded and 32 commits after the release Homebrew ships, and Ollama offers a cloud tag only. The sizing arithmetic, and the checks to run before you order hardware.
Read - AI News5 min
To Run Kolibri Locally, the License Is Apache 2.0 and the Runtime Is a Plugin Pinned Below vLLM 0.30
By Petru Popa ·Aleph Alpha released Kolibri, a 78B-total, 3.46B-active English-German MoE, under Apache 2.0 on 3 October. The model card answers what the announcement left open: the only supported runtime is Aleph Alpha's own vLLM plugin, which refuses vLLM 0.30; the 1M context is extrapolated past a 262,144-token config; and the chat template reasons at high effort unless told otherwise.
Read - AI News6 min
Ember-1 and LLM Cost Optimization: Fireworks' Own Table Adds Up to 23%, Not Half
By Petru Popa ·Ember-1 is Kimi K3 retrained to emit shorter reasoning, priced identically per token, so every saving is a token saving. The announcement says half the tokens. Summing the dollar column of Fireworks' own five-benchmark table gives roughly 23%, and the model is a serverless research preview with no weights, no fine-tuning and no on-demand option listed.
Read - AI News6 min
The 24 GB Card Fits the Weights, Not the Features
By Petru Popa ·Meta's Muse Glimmer targets consumer GPUs with 24 to 32 GB. Reading config.json and the GGUF file listing, the binding constraint on a 24 GB card is not the 131,072-token context — it is the vision projector and the speculative-decoding drafter, the two files that are the reason to pick this model.
Read