Run Kimi K3 using 29 GB of RAM at 0.50 tok/s

Kimi K3 — 2.78 trillion parameters — running on a consumer laptop.

$ waste run ~/models/k3.waste 'What is the capital of Italy?' waste: no --budget, using 46.24 GB of 64.00 GB (expert cache 17.56 GB) The capital of Italy is **Rome**. [16 tokens, 25.78 s, 0.62 tok/s | experts 9038 hit / 14514 miss = 38%] WASTE is an embeddable inference engine written in C, with no third-party runtime dependencies. It keeps the model trunk in memory, streams selected experts directly from disk, and uses the remaining RAM as a bounded expert cache.

Its current proof point is the complete open-weights Kimi K3 model: 2.78 trillion parameters, converted into a 982 GiB container and running on a 64 GB MacBook Pro at 0.45–0.62 tokens per second. This is not a distilled, pruned, or reduced variant .

WASTE was written for that one model and that one constraint: K3 does not fit in the RAM of current mainstream consumer systems. It is 1.42 TB as published and 982 GB after conversion. But a mixture of experts activates about 4% of itself per token, so almost all of that weight is idle at any instant — and idle weight does not need to be in memory, it needs to be reachable in time . WASTE keeps it on disk in a layout where one expert costs exactly one read, streams what each token actually needs, and spends every remaining byte of RAM on the part that repeats.

The engine is correct: every layer is validated against a PyTorch reference, the final logits agree to 3.6e-06, and the vision tower matches its own oracle to 2.3e-06. It is also slow — half a token per second, twenty-six seconds for the sentence above.

Both of those matter, and the second one should not be read as a disclaimer. We are not aware of another published demonstration of a model this size streaming from disk on a consumer machine: we found none for trillion-scale NVMe streaming, and the best-documented 671B-class recipes assume a server with a terabyte of DDR5. That is a report of what our search turned up rather than a survey — this repository carries no bibliography and no comparison table, so read it as an invitation to send a counter-example, not as a result. The interesting part is not the speed, it is that the whole thing is in the reachable range on a single consumer machine — and that from here the question is engineering rather than feasibility.

Where the levers were is not where they are. The two that looked biggest — reading fewer bytes per token, and keeping more of them in RAM — were both measured and both refused: this family's router has no tail to demote, and a cache the machine will not leave resident cannot be bought at any price. What paid instead was never about which bytes to read but when . Overlapping the expert reads with the arithmetic is worth ~1.6x; starting the next layer's reads on its own router's guess, one residual early, takes the hit rate from 14% to 38% at no extra bytes at all.

Both of those are exact — the cache statistics and the logits are unchanged — which is the property that makes them shippable rather than tuning. docs/EFFICIENCY.md is the account of how each lever was priced, including the three that were built before being measured and the two that were then taken back out.

What that opens up, concretely: a frontier-scale model that answers with no network, no per-token invoice, and nothing leaving the machine — which is the difference between "you may not send that data to an API" and "run it here". The format and the engine are not K3-specific in any deep way; K3 is simply the hardest case that exists today, and a model that streams at 2.78T streams comfortably at 48B.

Every number in this document was measured on the commit it is published with, and the ones that were wrong are recorded as wrong in docs/LEARNED.md rather than quietly corrected.

Every token answered by a cloud service is paid for twice: once on the invoice, and once in the electricity of a datacenter running a model that would fit — barely, awkwardly, but genuinely — on hardware already sitting on a desk. WASTE means to be the first concrete step toward ending that waste of tokens. The acronym came second.

disk, for the model 982 GB for the converted container — plan a terabyte disk, to convert it another 1.42 TB of staging for the published shards, freed afterwards RAM 29.05 GB minimum to open K3 at 4K context; 64 GB for the numbers here storage speed the container must be on internal NVMe — see below build a C11 compiler and make . No BLAS, no CUDA, no Python at run time Sizes here are powers of two, the way df and the engine both report them: the container is 982 GiB, which a disk vendor would call 1.05 TB.

The RAM floor is what the engine refuses to start below, and it is almost entirely the 27.28 GB resident trunk. Useful throughput starts higher: on a 64 GB machine the engine gives itself a 46 GB budget, of which 17.56 GB is expert cache, and that is the top of the measured curve. A 32 GB machine can technically open the model and will page badly; treat 64 GB as the real requirement.

Storage speed is not a detail. A token reads 17 GB of experts. On the internal SSD that is 12.78 GB/s and the model streams; over a USB enclosure it is 0.94 GB/s and the same token takes thirteen seconds. Convert onto internal NVMe, and use the external disk for the download only.

If a terabyte is not available, the same engine and the same format run Kimi-Linear-48B-A3B-Instruct from a 19 GB container with a 1.87 GB floor, at 10.7 tok/s. That is the good path for trying WASTE out before committing a disk to K3.

Self-contained. One libwaste.a , one waste binary, nothing at run time beyond libc and pthreads. Zero dependencies. No BLAS, no ONNX, no Python in the inference path, nothing to install. The Python under tools/ converts models and validates the engine; it never runs alongside it. Fully embeddable. Twenty-six public functions in src/waste.h : open a model under a RAM ceiling, generate, save the session, close. The CLI is a client of that API and touches nothing private — if the CLI can do it, so can an embedding host. waste_cfg cfg ; waste_cfg_init ( & cfg ); cfg . ram_budget_bytes = 46ULL << 30 ; /* a hard ceiling, not a hint; 0 sizes it to this machine */ waste_ctx * ctx ; if ( waste_open ( "/path/to/k3.waste" , & cfg , & ctx ) != WASTE_OK ) return 1 ; waste_generate ( ctx , ids , n , & params , on_token , user ); waste_close ( ctx ); The path is the container directory the converter wrote — no ~ expansion here, that is the shell's job.

Placement decides the speed A model is converted once into a .waste container: a JSON manifest, a resident trunk, and one expert bank per layer. Each expert record is 4 KiB-aligned with its gate, up and down matrices adjacent, so routing to an expert costs exactly one pread — not three, not a seek per matrix. The arithmetic was never the bottleneck.

Reads bypass the page cache ( F_NOCACHE on macOS, O_DIRECT on Linux, FILE_FLAG_NO_BUFFERING on Windows). That is deliberate: with a container smaller than RAM the kernel would cache everything, and the hit rates measured that way are a fiction that does not survive contact with a 982 GB model.

Every record's header is checked on the way in — right magic, the expert the index asked for, offsets that fit — so a bank that has been truncated or spliced stops the generation and names the record instead of answering from the wrong bytes. That costs nothing measurable. The record also carries a crc32 over its payload, and checking that is --verify , off by default: it is a pass over every record on every cache miss, about 5% on Kimi-Linear and 1% on K3. Worth it for a container you copied or downloaded and have not read since; not worth it on every token of one you converted yourself. See docs/FORMAT.md .

A layer knows all sixteen of its expert ids the moment its router runs, so the reads go out on their own threads and the arithmetic consumes them as they land instead of blocking on each. That is worth ~1.6x on K3, and the cache statistics are identical to the digit with it on and off — the engine does the same work, it just stops waiting for it.

The next layer's ids are not known: its router eats a hidden state that does not exist yet. But the router does exist, and it is resident. So at the end of each layer, once its own reads are consumed and the disk is about to go idle through the next layer's attention, the engine runs layer L+1's router on layer L's hidden state and starts fetching the six experts it names. One residual early, that guess is right 92% of the time at rank 1 and 81% over the first six.

It is exact by construction: the real router still decides, and the guess only decides when bytes move. The demand hit rate goes from 14% to 38% and the total bytes read do not change — those records were going to be read anyway. WASTE_LOOKAHEAD=0 turns it off.

The same trick in the prefill path was built and removed. A decode layer claims 16 cache slots so the speculative records survive; a chunk layer claims about 550, evicts them before use, and reads them twice — 6.9% more bytes for no time saved. docs/LEARNED.md §34–36.

Experts are stored as residual vector quantization — three stages of 256-entry codebooks over 8-dimensional vectors, 3.00 bits per weight — and the matrix is never materialized. For each token the engine builds a table of partial dot products, one per codebook entry per vector position, after which every expert row is three table reads and two adds.

The trunk stays at 4 and 8 bits. The model was trained with quantization-aware training on the experts only, so it has no trained tolerance for a squeezed trunk: a 3-bit trunk was built and measured, the cache prediction held, the throughput did not, and the output collapsed.

The most predictive number in this project. K3 touches 16 experts in each of 92 layers per token: 17.0 GB . Below that, an expert cached for one token is evicted before the next token asks for it, and the hit rate is not low — it is zero.

What crossing it buys has changed, though, and the table below is the first one to show it. Going from a 0% hit rate to 17% is worth about 8% of throughput now — 0.50 to 0.54 — because read-ahead already hides most of the I/O the cache would have saved. The sharp bend in this curve is no longer the climb above the floor; it is the collapse above 46 GB , where the engine stops fitting in the machine.

The hit-rate column predates the router lookahead, which roughly doubles it at every budget — 14% to 38% at 46 GB. It does not move the decode column's shape, because what collapses the 52 and 58 GB rows is the machine running out of memory and not the cache missing.

Ranges, not measurements , and the width is the finding. Every run behind a row reports cache statistics identical to the digit — the engine does the same work each time — so what varies is the machine, not the engine.

The two rows that fit are tight. 32 and 46 GB reproduce to within a few percent, because the engine's whole footprint fits with room to spare and nothing has to be taken from anything.

52 GB has no value. Two runs of the default configuration gave 0.04 and 0.15; three more with the trunk wired gave 0.46, 0.19 and 0.03 — seven-fold in the column above and fifteen-fold across both configurations, against 3652 hit / 8124 miss every single time. That budget sits exactly where the engine's footprint either does or does not fit beside whatever else the machine is holding, and which side it lands on is decided before the process starts. A row that spans 15x is not a slow row; it is a row whose mean would invite a comparison there is nothing to compare.

58 GB is uniformly bad and reproducibly so.

Order still matters, and more than the table shows. Re-run after the 52 and 58 GB rows have driven the machine into paging, 46 GB collapses — 0.02 tok/s in one such run — while again reporting identical counts. Sweep upward, one budget per quiet machine, and treat anything measured after a