We shrank Whisper large-v3 by 68%, below the size of turbo – Tileward

Tileward says it shrank Whisper large-v3 by 68%, bringing the speech-recognition model below the size of Whisper turbo, as part of a platform that compresses model weights and trims the context sent to inference. The company reports a 55% answer rate on six multi-agent problems against 63–67% with full transcripts, judged by a separate model, and says 50 of 2,251 probes in its own red-team test were blocked while 5 of 50 benign probes were wrongly blocked. Tileward prices model usage pay-as-you-go from $0.12 per million input tokens and offers deployment in a customer VPC or fully air-gapped.

Models + Context Serve compressed model weights on hardware you control, and send only the parts of a long conversation that can answer the next question. Models https://tileward.com/models/ Context https://tileward.com/context/ Every request costs less and stays inside your rules. Works with the clients you already run 28 have a recipe to paste https://tileward.com/docs/ connect , and anything else that speaks MCP works too. SOC 2 Type II and ISO 27001: in progress https://tileward.com/security/ certifications AI systems resend old context, serve more model than each task needs, and leave policy enforcement to prompts. That drives up inference cost and makes control difficult to verify. Ground answers in approved files with citations, and refuse restricted topics before the request reaches the model. Tileward changes the amount of model, context, and restricted work that reaches inference. Smaller model weights and shorter prompts return memory, context capacity, and serving headroom. Stop paying every agent and conversation to resend history the next question does not need. Make the policy decision before inference, then retain the verdict. Tileward is building the layer between AI applications and models: one place to reduce what reaches inference, define what it may use, and retain the decision record — across providers, VPCs, and air-gapped deployments. Each result links to the workload and method behind the number. Across six multi-agent problems against broadcast context, with ingest included. See the full results → 55% answer rate against 63–67% with full transcripts, judged by a separate model. Read the full comparison → 50 of 2,251 in our own red-team; 5 of 50 benign probes were wrongly blocked. See the red-team results → Model usage is pay as you go, from $0.12 per million input tokens. Every account includes Models, Context, Documents and Governance. No card and no time limit. Each model at its own rate, from account credit. Your VPC or fully air-gapped. We’ll distill the first and lock the second, try to get past both, and show you the record. Thirty minutes, an engineer, no deck. Pick a time right here on Google Calendar no form, no waiting on a reply . Prefer email? hello@tileward.com mailto:hello@tileward.com