Last Week in Chutes | August 19 to August 25, 2026
Burhan Khan
Engineering and product team.
Last Week in Chutes
August 19 to August 25, 2026
Week fifteen. We said out loud where the network is headed (a federation that outlasts any one company, including us), DeepSeek V4 Flash took the model spotlight with a 1M-token context, and Jon showed the napkin math for training on DGX Sparks.
The thousand-year thread
const put it bluntly: others are building AI companies, we are building an Intelligence Federation meant to last a thousand years. We quoted him and laid the vision out in full. Chutes.ai is not the network. It is one service provider on the network, and to last, the network has to evolve beyond one operating company. The thread describes a federation of service providers on shared inference infrastructure, one in the US, one in the UK, one in Singapore, each complying with its own local rules, all running on the same GPUs with the same end-to-end encryption. Underneath it sits conviction lock: locked stake becomes proportional, delegable access to the network's compute capacity, and conviction builds on chain with a time-weighted score that takes a long time to unwind. If a model gets restricted in one provider's jurisdiction, that provider stops serving it and delegates its access to a provider who lawfully can. The network keeps serving intelligence either way.
The thread: x.com/chutes_ai/status/2091888285748449320, posted 2026-08-24. More Chutes.ai updates are coming; the thread says more will follow.
DeepSeek V4 Flash, in the spotlight
The model spotlight this week went to DeepSeek V4 Flash: $0.44 per million input tokens, $1.32 out, and a 1M-token context window (prices are our 2026-08-25 snapshot; current numbers always at chutes.ai). Like every model we serve, it runs inside a hardware-attested TEE with end-to-end encryption. Nobody reads your prompts on the way through.
try it out: https://chutes.ai/app/chute/chutes-deepseek-ai-deepseek-v4-flash-0731-tee?tab=stats
Jon's Spark math
Jon walked through why ternary weights and NVIDIA's sparsity trick fit together. Ternary (-1, 0, 1) is naturally sparse, so forcing the 2:4 pattern that sparsity-capable NVIDIA GPUs reward costs roughly nothing in quality, nearly doubles expert throughput, and packs weights at 1.15 bits per parameter instead of 1.58. Then the hardware: a DGX Spark delivers about a petaflop of sparse FP4 compute with 128 GB of unified memory, a dual-Spark setup runs on a single 15-amp circuit, and his per-watt napkin estimate lands around 10x the training efficiency per wall-watt of a conventional GPU rig once you stop moving memory you do not need to move. All of it is design math, nothing measured on Spark hardware yet, and that is the point: it sketches a few thousand Sparks dispersed worldwide, each sipping a few Mbps. That is what Parallax enables.
Jon's post: x.com/jon_durbin/status/2090125585519444276
Muse Glimmer 30B joined the catalog
Meta's Muse Glimmer 30B went live at $0.0245 per million input tokens and $0.0978 out, which ties the cheapest tier we serve, with images and text in and a 131K context (our 2026-08-20 launch snapshot; current numbers at chutes.ai). Point any OpenAI-compatible client at llm.chutes.ai/v1 and you are running it. Inside a hardware-attested TEE, as always.
Dev Corner
This week's theme: getting a new developer from zero to a first call, faster.
The Starter Guide moved into the docs, with live pricing. The new guide at chutes.ai/docs/guides/starter-guide takes you start to finish, account creation to a live API call in one sitting, and now shows live model prices instead of a stale table. Any OpenAI-compatible client works; swap the base URL to llm.chutes.ai/v1.
What we're working on next
Offline, reproducible attestation measurements. We can now generate a confidential VM's full attestation measurements on any ordinary host, not just the exact hardware it runs on. Initial validation is done; verification and documentation are being finished now. This is the foundation for letting anyone independently verify what is running, and it also simplifies rolling out VM updates and coordinating them with miners.
The chutes-cvm operator toolkit. A single installable command-line tool that sets up, verifies, and generates measurements for a confidential VM. It is what makes offline measurement practical for miners and third parties.
Next week
More Chutes.ai updates from the federation thread.
Verification work keeps moving toward independently checkable measurements.