Telperion released two modified versions of Qwen3.8-27B, Qwen3.8-27B-MLA and Qwen3.8-27B-GLA-g2, which take the attention mechanism the model shipped with and swap it out: grouped-query attention gone, DeepSeek-style multi-head latent attention and gated linear attention in its place, weights on Hugging Face. The claim is roughly a 4x reduction in KV cache with no retraining of the base model. For anyone serving long context on fixed VRAM, the KV cache is the wall that arrives before parameter count does, so a 4x cut is the difference between holding 128k and holding something closer to 500k on a single card. The benchmarks are the author's own, and retrofits trade something, so the quality question stays open until the variants meet somebody else's eval set.
What Telperion did to the model is a smaller story than when they did it. Attention architecture has been a pretraining decision, locked at the moment a lab commits compute. Here it became a post-release modification, applied by a party with no relationship to the original training run. Qwen3.8-27B picked up four separate mentions over the week, and the character of them shifted off whether it runs and onto what people are doing to it. A Zellige AI finetune goes after token efficiency for CJK text, which compounds across every request a team makes. Somebody ported the model to a PlayStation 5 GPU. A model becomes a substrate at the point third parties modify it faster than its authors ship revisions.
The same post-release logic showed up twice in hardware terms. A 35B-A3B build of MiniCPM-V 4.7 is running vision and chat on a handset in roughly 3GB of memory, sparse-active-parameter routing landing on mobile silicon instead of in a paper. NVIDIA has Nemotron Bolt 3.5 at 74B total and 9B active dated for today. Nine billion active puts it in the same deployment envelope as the 27B to 32B dense models most small teams run as their default, at more than double the total size. Active-parameter count is now the number that describes the hardware requirement, and total parameter count has become closer to a labeling convention. The operator question is which of the two a team's capacity plan is written against, because one of them is off by a factor of eight.
On the readability side, Simon Willison documented indie developer Popcar's account of AI-assisted decompilation moving from stunt to volume, producing buildable source from shipped game binaries. The recovery case is the pleasant one: code whose source is long gone is reachable again. The other edge is that a shipped binary is a more legible document than most teams priced in on the day they shipped it. Set next to the Telperion retrofit, the two describe one condition. A compiled binary and a trained checkpoint are both editable artifacts in the hands of whoever downloads them. We argued on Monday that replication stays the exception; the week leaned the other way, and last Friday's generative-UI case, reproduced inside a day, was the leading edge of it.
Two tool releases went after the bindings around build systems. xtool builds and deploys iOS applications from Linux and Windows using SwiftPM, which takes the macOS runner out of an iOS pipeline and, more usefully, lets an agent on a Linux box compile and ship an iOS target instead of writing Swift it can never execute. OpenCode 2.0 shipped as an agent harness with inspectable internals and an arbitrary model endpoint, local ones included. Both move a capability out from under a vendor's toolchain. An agent that cannot run the compiler is a text generator with opinions about Swift.
The commit-volume thread picked up real substance. A teardown at sunshowers.io traces Scion, Anthropic's Git replacement, to Mononoke, the Mercurial-based monorepo server Meta has operated for about a decade. The lineage is the useful part: the answer to agent-rate commit volume is a known architecture with ten years of operational history behind it, which is a considerably better position than inventing one. Anyone who followed the Mercurial and Git arguments in the trade press as they were happening will recognize the shape of the fix. Tuesday's piece took the unreviewed daemon from the governance side; Scion is that same volume pressure surfacing in storage.
There is a cost inside all of this that the retrofit framing tucks away. Jensen Huang, CEO of NVIDIA, put the general version plainly: "You sell a chip one time, but when you build software, you maintain it forever." A latent-attention retrofit of someone else's checkpoint is a fork, and a fork has an owner. When Qwen ships its next revision, the MLA variant either gets rebuilt against it or quietly ages out underneath a production serving stack that still points at it. I would want that answered before standardizing on a retrofit, and the answer is a staffing answer, not a benchmark.
Three numbers carry the week: 4x off the KV cache without touching the training run, 3GB for a 35B multimodal model on a phone, 9B active inside 74B total. All three describe the same movement from different angles. The deployment envelope a team planned against six months ago has shifted, and it shifted through modifications made after the artifacts shipped, by people who did not train them.