by Chris DePuy, 650 Group / October 14, 2026
The most interesting inference demo of the past week did not serve more tokens per second. It learned. TensorFold 1.0.3 shipped Living Weights on October 9 — a serving engine that writes a taught fact into the model’s own weights, where it survives a restart — and two hours and forty-five minutes later a community operator had filed the pull requests that carry the capability to GLM-5.3, announcing a port in which the full 753-billion-parameter model, split across four M3 Ultra Mac Studios, learned a fact in about 24 seconds and answered from it, including a phrasing it was never taught. The maintainer declined both pull requests the same morning, with thanks and with reasons. I think the sequence matters more than either side’s code: it shows how quickly the gap between the model a lab ships and the model an operator wants now gets closed, and it shows a maintainer using a declined contribution to state, precisely, where the design is going.
Living Weights: facts written into the weights, not the context
The project’s Living Weights guide is explicit about what the feature is not: “not a prompt cache, a retrieval index or an adapter.” Each kept fact takes the physical form of roughly ten new neurons persisted into the checkpoint’s own safetensors files, and none of the model’s existing weights move. The mechanism runs through the mixture-of-experts shared experts: every MoE layer carries 64 pocket neurons that stay zero until a fact uses them, the first save grows the shared expert’s tensor from 3,712 to 3,776 slots and records the widened shape in config.json, and a single fact claims one neuron in roughly ten of the model’s layers. Learned facts live in the model folder, keep working after a restart, and serve on a Mac or an NVIDIA box started from the same folder; there is no per-fact undo, so the guide’s standing instruction is to learn only on a copy and keep the pristine original elsewhere. The 1.0.3 release notes shipped the feature conservatively as Sliding Weights (experimental): learning ran on Apple silicon against Nemotron 3.5 Lightning in its MLX 4-bit build only, NVIDIA GPUs refused the flag though a CUDA build would serve a folder that learned on a Mac, a fact took two to four minutes, and on a five-fact test document 5 of 11 recall questions came back while 1 of 16 neighbouring questions leaked a fact.
The port: GLM-5.3 learns a fact in 24 seconds across four Macs
The port is documented in two pull requests filed at 02:32 UTC on October 10, each closing with a note that it was generated with Claude Code, and the announcement followed fifteen minutes later — the operator’s own words for the labor are “me and my coding agent.” The full-model pull request carries the measured section that matters: with no change loaded, its replies equal the stock build exactly across a 391-prompt reference set at 53.9 tok/s; “I like blue.” learned on four ranks in 24 seconds with a 20-second commit, the saved shard byte-identical on all four ranks, the KV caches untouched on every step and replies unchanged before the commit; after the commit, the trained question and an unseen phrasing both answer blue, and six of seven neighbouring questions are bit-identical — the seventh, who wrote Romeo and Juliet, was reworded and still correct. The pull request also documents its own catch: an early run reused an exchange sequence number and the ranks saved different shards, so the branch now refuses a reused number. The operator put money behind it minutes later: “I just supported @ashxhart with $300 on @buymeacoffee! ” he wrote. The single-Mac lane is the second pull request: GLM-5.3-Flash — a 180.6 GB checkpoint — loads on one 256 GB Mac behind an opt-in load limit, runs token-exact against stock with no shard (20 of 20 greedy replies equal, last-row logits byte-identical, decode 49.0 to 49.4 tok/s with the shard against 48.7 to 49.2 without), and learned “I like blue.” end to end in about 7.5 minutes. That pull request’s known-limits section is blunt — “The guard has not yet let a GLM fact stay,” because a held-out “which colour do you like best?” also answered “You like blue best” — which makes the four-Mac 753B result, with its strict recall check passing on a third-person fact after 20 steps, the stronger receipt of the two.
The maintainer’s answer: thanked, declined, and pointed at a different design
Both pull requests were closed unmerged at 10:03 UTC the same morning, seven and a half hours after filing. On the Flash port, the maintainer wrote that getting a lesson to run end to end on GLM-5.3-Flash on one Mac, with the engine token-exact when no shard is loaded, “took real work,” and called the turn-end fix for chat templates that never close an assistant turn “a good catch” — but the feature is moving to changes written into the model’s own shards, and side-file shards with an index are “the design we’re leaving,” with about half of the ported learner duplicating the Nemotron learner line for line at the moment a rewrite is making that learner generic across backends. On the full-model port, learning in lockstep across four Macs with byte-identical shards is “neat work” and the reused-sequence-number catch “was a good find,” but the pull request starts from an unmerged server refactor unchanged — an 890-line server of its own, a copied core module, a 1,457-line forward pass and a 2,319-line kernels file, plus four new Python tools while the project is “moving the project off Python rather than adding to it.” I read the pair of closures as the healthy version of the not-invented-here reflex: the demonstrated capability stayed on the roadmap, the two good catches entered the record, and roughly 4,700 lines of parallel implementation did not enter the tree.
The releases since: Living Weights on the GB10, faster and more reliable
Six releases landed in the four days after the port, and each moves Living Weights up the ladder on hardware the desk market actually buys. A sixth release, 1.0.5, rounded out the ladder the same UTC day with reliability work: Living Weights learned from Japanese and other text written without spaces, held a fact told in one language against prompts in that language, and its leak check stopped counting an everyday phrase — “If you’d like, I can help” — as a leak that undid an “I like blue.” lesson. 1.0.6 brought learning to the GB10 — the DGX Spark — where the native CUDA server runs the same learner, lessons, checks and saved shards as the Mac, and folders taught on either backend serve on the other; the guide now holds back roughly 26 GiB on the GPU for the trainer when learning runs on a GB10, and prices a fact at 2.5 to 3 minutes there against 1.5 to 2.5 on an M5 or M3 Ultra, and 1.0.6’s notes carry the cost figures — a decoded token on a GB10 running about 1% more with no lessons, 6% with one kept lesson and 11% with four. 1.0.7 made learning about four times faster, holding up to 32 of a lesson’s answers side by side through its training passes. 1.0.8 improved reliability and scope: ten facts told in turn bring back 18 of their 20 questions on an M5 Ultra against 12 in 1.0.7, and 13 to 16 on a GB10 against 10; a newer fact about a changed value replaces the earlier one; and facts now cover people, things and companies, not just the user. Its known-issues list is the honest residue — facts about the same subject overwrite each other, yes-or-no questions about a changed value are unreliable, facts pile up badly past a few dozen, and thinking on brings back fewer answers — alongside the cost line that an M5 Ultra decodes at 213 tok/s after ten facts and 205 after a hundred, against 248 before any. 1.0.9 fixed GLM-5.3-Flash’s sparse-attention and Living-Weights-training kernels on M1 Macs, where a wrong thread count had left attention outputs at zero and could leave gradients unwritten with no error — a GLM-side kernel fix that stops short of the feature gate, which still admits only Nemotron. And 1.0.10, published overnight, turned the cost curve around on the GB10: layers whose pocket holds no neuron in use no longer get paid for, so a copy that learned one fact decoded 17% faster than the prior release with drafts off, 5% to 10% faster drafted, and read 2K to 32K-token prompts 8.5% to 10.3% faster, with the same replies.
The open gap, and what the week says about the market
The capability the port demonstrated has not landed upstream. The guide, fetched this morning, still names Nemotron 3.5 Lightning in its MLX 4-bit build as “the only model it learns into today,” and 1.0.8’s behavior changes have every Mac model but Nemotron refuse the learning flag at startup, since Living Weights learns into Nemotron only. My caveat on the whole chain: every figure here is a single-maintainer or single-operator measurement — the 24-second lesson and the 753B demo are the port author’s own runs on his own four Mac Studios, and the release benchmarks are the TensorFold project’s own hardware — pending the wider replication this tier of the market always waits for. The signal is still legible. Learning in weights turns a model purchase into something closer to a persistent system: the fact survives the restart, travels with the model folder to other machines, and costs about ten neurons instead of a growing context bill, and the long-document mode remains the weak spot the guide itself reports, with a 21-fact company handbook keeping about half its facts and answering 6 of 16 questions. It is also where the deskside tier moves first, because both the four-Mac Studios demo and the GB10 trainer sit at prices a department can buy — and because the feature’s whole value, a model that stays taught, only matters on hardware someone keeps. As firms such as 650 Group have been tracking in their AI infrastructure research, the deskside and workgroup tiers keep converting capabilities the large labs treat as research projects into product surface within days of the first public demonstration, and this week’s lesson is that the gap between what a model ships with and what an operator needs is now measured in hours, not release cycles.