Mistral AI announced the release of Mistral Large 4 (ML4), internally nicknamed Le Chonk, as a public preview on October 6, 2026—a granular Mixture of Experts model with 1.05 trillion total parameters, 49 billion active per token, a 1.6 billion parameter vision encoder, and a 1 million token context window. The API is live now at $1.36 per 1M input and $4.18 per 1M output tokens, but the weights do not ship until end of October, so self hosting is not yet possible. MarkTechPost, Kingy.ai, and Pasquale Pillitteri covered the release. The model shipped. The weights did not.
I run ten production platforms. The construction ERP at coenconstruction.com calls GPT-4 to parse change-order language. The estimating SaaS at estimate.pro calls Claude to extract line items from supplier PDFs. Curbscript calls Gemini to reconcile scheduling conflicts in punch-list data. Every one of those integrations went through the same evaluation loop before I shipped: call the API on test data, measure accuracy and latency, check the cost per call, compare output quality across three sample workflows, run it for a week on staging traffic, then flip the environment variable in production. I cannot do any of that with Mistral Large 4 because the only artifact that shipped on October 6 was a callable endpoint. The limitation is important: Reflection had not released Beam's weights, technical report, model card, or developer artifacts as of October 7, 2026. You cannot download it yet—as of 6 Oct 2026 there are no weights to download, and the launch post says Beam is undergoing final red-teaming and evaluations, with Reflection releasing the weights, technical report, model card and developer artifacts later this month. Two labs shipped models this week. Neither lab shipped the files.
The API is live but I cannot test the model offline
The dog-training platform at valhalla-k9.com stores client training schedules. The freelancer site at hiremariaelena.com handles intake forms and payment records. The permit-sketching tool processes zoning compliance summaries. Every platform calls a model. Every model integration started with an offline evaluation that ran locally or in a private VPC. I load a checkpoint, point it at a CSV of sample inputs, run inference on two hundred rows, measure the output, compare it to the previous model, and decide whether to route production traffic. What has not arrived is the part the words open-weight describe—there are no checkpoints to download today, no licence file, and no repository. Mistral published an API. I cannot download the model the API calls.
The construction ERP generates change-order summaries that subcontractors sign. The estimating SaaS produces bid packages that go to municipal procurement offices. The Punchlist at thepunchlist.vercel.app submits zoning-compliance reports to city portals. The model can ship but the capability I need does not always arrive with it. The capability I need this week is offline evaluation. The announced open-weight release is scheduled for the end of October; Mistral's Hugging Face upcoming-release page now lists October 31 as the current expected date. The API preview launched October 6. The weights ship twenty-five days later. I can call the model in production today. I cannot test the model on my own infrastructure until the end of the month.
Reflection AI announced a 501-billion-parameter model with the same gap
Reflection AI launched Beam, its first open-weight model, on Monday, positioning the 501-billion-parameter system as a coding and reasoning tool for enterprises that can match Chinese open models at lower inference cost. Weights are due later in October under an Apache 2.0 license, and no independent evaluation exists yet. Quartz, Technology.org, and Superpower Daily covered the announcement. Reflection opened a waitlist. The waitlist is not a download. The model uses a sparse mixture-of-experts architecture. The model uses a sparse mixture-of-experts architecture, activating 23 billion of its 501 billion parameters at inference time. I cannot evaluate inference cost until I can run inference.
The permit tool parses zoning ordinances and generates compliance checklists. The Twilio dialer at the construction ERP routes outbound calls to subcontractors. The invoice workflow at the estimating SaaS posts approved line items to QuickBooks every night. None of those workflows will call Mistral Large 4 or Reflection Beam until I can test the models offline. The API pricing is temporary but the integration is not. The integration requires an evaluation. The evaluation requires a checkpoint. The checkpoint does not ship for three more weeks.
The API returns 200. The weights return a 404 and a calendar invite.
Mistral trained Large 4 on 3,800 Nvidia Grace Blackwell GPUs in European data centers. Reflection trained Beam on 10,500 Nvidia GB300 GPUs. Both labs published benchmarks. Neither lab published the files that would let me verify the benchmarks. The platform markets to enterprises but the API does not ship with the tooling enterprises need. The tooling I need is a downloadable checkpoint and a license file. The API preview went live October 6. The checkpoint arrives October 31. Until then the model is callable in production but not testable offline. The labs call that an open-weight release. I call it a three-week lag between the launch announcement and the artifact I can actually evaluate.