Kimi K3 — 2.78 trillion parameters — on a 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, 49.31 s, 0.32 tok/s | experts 3357 hit / 20195 miss = 14%]
That is the full open-weights K3 — 93 layers, 896 experts in each of the 92 that are MoE, multimodal — answering on a MacBook Pro, from a 982 GB container on the internal SSD, inside 46 GB of RAM. Not a distillation, not a pruned variant, not a smaller sibling. The model Moonshot published, with its own tokenizer, its own chat format and its own vision tower, running on hardware you can buy in a shop.
WASTE was written for that one model and that one constraint: K3 does not fit in RAM, and no consumer machine will ever hold it. 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 — a third of a token per second, fifty seconds for the sentence above.
Both of those matter, and the second one should not be read as a disclaimer. As far as we could find, nothing else runs a model of this size from disk on a consumer machine; the literature we surveyed has no demonstration of trillion-scale NVMe streaming at all, and the best-documented 671B-class recipes assume a server with a terabyte of DDR5. The interesting result is not the speed, it is that the whole thing is in the reachable range on hardware that costs less than a month of API bills — and that from here the question is engineering rather than feasibility. Half of a decode step is already disk I/O running near the drive's measured ceiling, so the levers are known: read fewer bytes per token, and keep more of them in RAM.
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
1.86 GB floor, at 8.92 tok/s. That is the good path for trying WASTE
out before committing a disk to K3.
- Self-contained. One
libwaste.a, onewastebinary, 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, ¶ms, on_token, user);
waste_close(ctx);The path is the container directory the converter wrote — no ~
expansion here, that is the shell's job.
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.
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. Above it the curve bends sharply.
| budget | expert cache | hit rate | decode |
|---|---|---|---|
| 32 GB | 3.32 GB | 0% | 0.31 tok/s |
| 46 GB | 17.32 GB | 13% | 0.32 tok/s |
| 52 GB | 23.32 GB | 27% | 0.11–0.14 tok/s |
| 58 GB | 29.32 GB | 37% | 0.04 tok/s |
Measured in that order, on an otherwise idle machine. Order matters: re-run after the 52 and 58 GB rows have driven the machine into paging, 46 GB gives 0.22–0.25 rather than 0.32 — while reporting hit and miss counts identical to the digit. The engine is deterministic; the machine is not, and it does not fully recover between runs. Sweep upward.
Everything in the memory design exists to get above that line, which is why the engine works to free RAM rather than to save it.
And there is a ceiling on the other side, closer than it looks. Read that table twice: the hit rate climbs all the way down. At 58 GB on a 64 GB machine the cache serves 37% of experts from RAM and the engine is eight times slower than at 46 GB, where it serves 13%. The engine is inside its budget; the machine is not, so the OS pages out the expert cache, and a "hit" becomes a page fault instead of the disk read the engine was managing.
So the usable window is narrow. It opens at ~46 GB, where the cache finally clears one token's working set, and it has already closed by 52 — on an otherwise idle machine, with 49 GB free before the run. It is also sharp enough to move under a change that looks unrelated: taking 1.11 GB of embedding table off the resident set fed straight into the cache at a fixed budget, and that was enough to push 58 GB from 0.32 tok/s to 0.04.
So the default does not fill the machine. Expert cache is only worth anything in whole multiples of that working set, and the remainder above a multiple buys a few points of hit rate while pushing the machine towards paging. When it picks a budget for itself the engine steps down a whole working set at a time and takes the largest that fits under seven eighths of RAM: K3 asks for floor + 3× — 80.63 GB — and gets floor + 1× on this laptop, a 46 GB budget and a 17.56 GB cache. That is the top of the curve above, reached with no flag. A 128 GB machine still gets the full 3×.
An earlier version took every byte up to the cap instead, which put a 27 GB cache on this machine — between two budgets measured at 0.11 and 0.04 tok/s. The real lesson is that a cache you do not control is not a cache, and the corollary is that an engine should stop asking for memory before the OS starts taking it back.
K3's attention is a 3:1 hybrid: Kimi Delta Attention, which carries a
fixed-size recurrent state instead of a growing KV cache, and gated
multi-head latent attention. The MLA layers cache the 512-wide latent
rather than expanded per-head keys and values, with kv_b_proj absorbed
into the query and the output:
q_nope · (W_kb c) == (W_kbᵀ q_nope) · c
Σ_s a_s (W_vb c_s) == W_vb (Σ_s a_s c_s)
Identical logits to 1.2e-05, and 53× less cache: 11.25 GB becomes 0.21 GB at 4K context. It is also what makes long context possible at all — the expanded layout wants 360 GB at 128K tokens, the latent one 7.2.
MacBook Pro M5 Pro, 64 GB, container on the internal SSD. Every figure was measured on the commit it is published with.
| minimum RAM | 29.05 GB at 4K context |
| 30.54 GB at 32K, 35.63 GB at 128K, 83.21 GB at 1M | |
| resident trunk | 27.28 GB |
| read per token | 17.0 GB at ~9.9 GB/s, near the SSD's measured ceiling |
| model load | 20 s |
| prefill | 0.47 tok/s chunked, 0.29 sequential |
| decode | 0.32–0.34 tok/s at the default budget, the best this machine gives |
| vision tower | 15.7 s for a 1024-patch image, 27 layers |
| image in a prompt | 256 positions for 896x896, 2.8 s each — as text |
The floor is almost entirely the resident trunk. Useful throughput starts above ~46 GB, where the expert cache finally clears one token's working set, and is gone again by 52, where the machine starts paging. Below the first line extra RAM buys nothing; above the second it costs, badly. The window is one budget wide on this machine.
The tower is not what an image costs. Encoding 1024 patches takes 15.7 s;
the 256 positions it produces then go through the 92 MoE layers like any
other token, which is the other 731 s. An image is priced as text of the same
length, so the patch budget in vision.json is a real dial: halving the
grid halves the prompt.
| minimum RAM | 1.86 GB |
| decode | 8.92 tok/s at an 8 GB budget, 78% cache hit |
The same engine and the same format, on a model that fits comfortably. This is what WASTE looks like when it is not fighting.
Decode on K3, 17.32 GB of cache and still cold — 6.7% hit over ten steps, which is the state a fresh prompt starts in:
| share | |
|---|---|
| MoE, all of it | 82.5% |
| of which expert I/O | 53.5% |
| of which expert matmul | 20.0% |
| KDA layers | 14.5% |
| MLA layers | 2.8% |
| lm_head | 0.2% |
Reproduce with WASTE_PROFILE=1 WASTE_CACHE_MB=17735 ./test_forward MODEL 1008,10484,318,15383,387 out.bin 5. The I/O share falls as the
cache warms, so a long session sits lower than this; the ranking does
not change.
The I/O already runs near the hardware limit — 17.0 GB per token at ~9.9 GB/s against the SSD's measured 12.78 — so it only gets cheaper by happening less often, which means cache, which means RAM. That is the whole optimization story so far, and the reason the next steps are about memory rather than arithmetic.
git clone https://github.com/sqliteai/waste && cd waste
make # libwaste.a, waste, libwastevq
make check # 23 pass, 11 skip on a fresh cloneNo configure step and no dependency resolution. make check needs no
model: it builds a small synthetic container and runs the engine against
it. The eleven skips are the checks that need something a clone does not
carry — the PyTorch oracle, the round-trip against the source shards,
anything driving the CLI with text, since the synthetic container carries
no tokenizer, and the K3 checks, which want the container and the release
on disk. With both containers present the suite is 36 checks.
Conversion is the one step that needs Python, and it happens once. The source is moonshotai/Kimi-K3 exactly as published — 96 safetensors shards, 1.42 TB, nothing patched:
# 1. preflight: reachable? how big? does it fit?
tools/fetch_weights.sh --dest /Volumes/staging/k3 --dry-run
# 2. download — resumable, safe to kill, safe to re-run
tools/fetch_weights.sh --dest /Volumes/staging/k3
# 3. convert into a container
uv run --with torch --with safetensors python tools/convert.py \
--src /Volumes/staging/k3 \
--out ~/models/k3.waste --jobs 3That produces the 982 GB container every number above was measured on. It takes about 4.7 hours with three processes on the M5 Pro (23.7 with the pure-torch encoder — see docs/K3.md), and wants ~1.0 TB free on the target volume. The converter is resumable too: a layer whose bank is already written is skipped, so an interrupted run costs only the layer it was in the middle of.
The download is the part that goes wrong. A 1.42 TB pull over hours will
hit dropped connections, CDN 5xx and at least one interrupted run, so
every shard resumes mid-file rather than restarting, retries with
exponential backoff and jitter, and counts as done only when its size
matches Content-Length — recorded in a state file, so a re-run skips
finished shards without even a HEAD request. --check re-verifies
everything on disk against the remote and downloads nothing (96 shards in
34 s). --repo points it at another model, HF_TOKEN at a gated one.
macOS and Linux.
Give --dest a staging disk rather than the volume that will hold the
container. The shards are read once, by the converter; the container is
read continuously, at every token. On this machine the external enclosure
measures 0.94 GB/s against the internal NVMe's 12.78 — see
docs/GATES.md, Gate H — which is the difference between
a model that streams and one that stalls.
tools/pipeline.sh chains the whole thing unattended — download, convert,
round-trip the container against the source weights, generate, then diff
the logits against the PyTorch oracle — and leaves a report next to the
container. The same converter handles the other member of the family,
Kimi-Linear-48B-A3B-Instruct, into the 19 GB container of the second
benchmark; --src is the only thing that changes.
Pre-converted containers are on their way to huggingface.co/sqliteai, at which point this whole section becomes a download and the Python is only needed for models we have not published.
The container is the directory the converter wrote, so give it that path —
~/models/k3.waste throughout this README:
waste run ~/models/k3.waste "The capital of France is" -n 32
waste chat ~/models/k3.waste # multi-turn, state kept
waste eval ~/models/k3.waste "2 + 2 =" --top-k 5 # next-token distribution
waste plan ~/models/k3.waste --budget 46G # what fits, what does not
echo "prompt" | waste run ~/models/k3.waste # stdin works too-n is a cap, not a requirement: without it generation stops at the
container's end-of-sequence token or at 128 tokens, whichever comes
first. The examples pass it because 128 tokens of K3 is six minutes.
--budget is optional, and leaving it out is the right default rather
than a fallback: the engine takes the container's recommendation, steps it
down a whole token working set at a time until it fits under seven eighths
of physical RAM, and never goes below the floor — a budget you set
explicitly under the floor is refused rather than swapped into. It then
says on stderr what it landed on, so the same command on two machines is
not silently two different runs:
waste: no --budget, using 46.24 GB of 64.00 GB (expert cache 17.56 GB)
--verify checks each expert record's crc32 as it comes off the disk.
It is off by default, and that is a throughput decision rather than a
claim that containers do not rot: it is a pass over every record on every
cache miss, about 5% on Kimi-Linear and about 1% on K3, where the read
dominates. Turn it on once for a container you copied, downloaded, or left
on a disk you do not trust, and for anything whose wrong answers would be
believed; leave it off for one you converted yourself and have been
reading since. WASTE_VERIFY=1 in the environment does the same thing,
and it is the only way in for the server, which has no flag of its own.
Either turns it on; neither turns it off.
What is checked either way: a short read, and a record header that does
not describe the expert the bank index asked for. Those are O(1), they
cost nothing measurable, and they are what keeps a damaged offset out of
the arithmetic — --verify only adds the pass over the payload.
waste --help lists all nine commands. --json makes eval, tokenize,
plan, info and bench machine-readable.
serve/ is an OpenAI-compatible HTTP server — the second client of the
public API, alongside the CLI, reaching the same engine through ctypes
rather than keeping a copy of the model code in Python:
make libwaste.dylib # or libwaste.so on Linux
python3 -m serve ~/models/k3.waste --port 8000curl localhost:8000/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{"model":"k3","messages":[{"role":"user","content":"Why is the sky blue?"}]}'/v1/chat/completions (streaming and not), /v1/completions,
/v1/models, /health. It carries the whole of K3's prompt format, not
the four-string subset a container's chat.json can hold: tool
definitions and tool results, typed call arguments, JSON response schemas,
tool_choice, the think channel and thinking_effort, and images — plus
the parser that reads the reply back into reasoning, answer and
tool_calls. Stdlib only.
The prompt renderer is a port of encoding_k3.py from the release, and the
test suite checks it against that file segment for segment on a corpus
of 38 conversations whenever the weights directory is on disk.
docs/SERVE.md is the reference.
K3 is multimodal — a 401M ViT, 27 layers, patch 14 — and so is the engine.
--image attaches a picture; repeat it for several:
$ waste run ~/models/k3.waste 'What is in this picture?' --image landscape.png
[landscape.png: 192 image tokens]
The picture shows a simple, stylized landscape with:
- A **blue sky** with a gradient from darker blue at the top to lighter blue near the horizon.
- A **yellow sun** in the upper right.
- A **gray hill or mountain** in the middle distance.
- A **green field** covering the lower part of the image.
[78 tokens, 234.25 s, 0.33 tok/s | experts 15314 hit / 99502 miss = 13%]
That is a 448×336 image, and every element of the description is in it — including the sky gradient, which is the kind of detail that separates a tower that works from one that merely runs. The picture was generated by a twenty-line script rather than photographed, so the answer can be checked against what was drawn instead of against an impression.
PNG, JPEG, GIF, BMP, TGA and PSD, decoded by the one vendored header in
third_party/. It works on run, chat and eval; inside a chat,
/image FILE attaches a picture to the next message, and it is spliced
once — the positions are in the attention state afterwards, so later turns
discuss the same photograph without re-encoding it. The 27-layer ViT is
loaded only when an image is present, because its 434 MB otherwise come
straight out of the expert cache.
An image is not one token. The tower turns a 14-pixel patch grid into one
embedding per merged 2×2 patch, and each occupies a position in the
sequence — the 448×336 above is 192 of them, an 896×896 photo at the
default budget is 256. That is worth knowing before wondering where a
context window went, and it is most of what an image costs: the 234 s in
the transcript is the 78 generated tokens alone, and the picture is paid
for before that, in prefill. An image is priced as text of the same
length. The tower is the cheap part — 15.7 s for a full 1024-patch
image — and its output then walks through 92 MoE layers like any other
token. Halving max_patches in vision.json halves the bill.
Through the library it is three calls, because a host needs to size the prompt before committing to it:
size_t rows;
waste_image_add(ctx, "photo.png", &rows); /* encode and queue */
waste_image_expand(ctx, raw, n, ids, cap, &n_ids); /* placeholder -> N */
waste_generate(ctx, ids, n_ids, ¶ms, cb, u); /* consumes the queue */The tower's shape, the patch budget and the pixel normalization live in
vision.json, which the converter writes from the release's own nested
vision_config and from preprocessor_config.json. K3 normalizes to
[-1, 1] with mean = std = 0.5.
That last sentence was wrong here for a day, and the way it was wrong is worth keeping. This section used to say K3 ships no preprocessor config, so the normalization was "the CLIP convention this lineage of towers uses rather than a value read out of the release" — an assumption, labelled as one. The release does ship the file; the downloader fetched a hardcoded list of filenames and never asked the repo what it contained. The tower still matched its oracle at 2.3e-06 throughout, because the oracle is fed random pixels and never touches the normalization. An honest caveat is not a substitute for reading the file.
| build | model-free suite | backend | |
|---|---|---|---|
| macOS arm64 | yes | 23 pass / 0 fail / 11 skip | NEON |
| Linux arm64 | yes | 23 pass / 0 fail / 11 skip | NEON |
| Linux x86_64 | yes | 23 pass / 0 fail / 11 skip | AVX2 |
| Windows x86_64 | yes | container, CLI and forward pass — see below | AVX2 |
The first three run the same suite and now agree check for check: same
23 passes, same 11 skips, same list. CI has no container, so
tests/run.sh builds a synthetic one and the checks that need real
weights say SKIP rather than passing quietly. All three also pass the
sanitizer suite and 400 fuzz cases.
Windows is cross-compiled with MinGW-w64 on a Linux runner and then run
on a Windows one: the binary reads a synthetic container, opens it from
the CLI, and produces the same logits token-by-token as it does in
chunks. It is not the same suite — tests/run.sh is a bash script that
rebuilds first, and the Windows job runs binaries it did not build — so
what is claimed is what that job checks and no more. Nobody has run it on
a real container there.
The platform is the variable, the suite is not, and that is the point: both Linux targets produce the same continuation as macOS and pass engine matches the PyTorch oracle when given a container, so the numerics carry across architectures and compilers.
SIMD is selected at run time from CPUID, so a single x86 binary uses AVX-512 where it exists and AVX2 where it does not. Accelerator backends are build-time options. A Metal backend exists and is off by default because it is correct and 22% slower: this engine issues several hundred small dependent matvecs per token, the worst possible shape for an accelerator, and the CPU path already runs at the machine's memory bandwidth.
src/ the engine — 6,700 lines of C, no dependencies
model.c forward pass, MoE routing, KDA and MLA layers
ecache.c bounded LFRU expert cache over the banks
vision.c the 27-layer ViT and the projector into text space
image.c a file on disk to the patch tensor the tower wants
waste.c the public API
simd_*.c per-ISA kernels, selected at run time
cli/ the CLI, a client of the public API
serve/ the OpenAI-compatible server, the other client
xtml.py K3's prompt format, ported from the release's encoding_k3.py
regions.py its replies, back into reasoning / answer / tool calls
engine.py libwaste through ctypes, and the request queue
server.py /v1/chat/completions and friends
tools/ conversion and validation (Python, never at run time)
docs/ format, engine, backends, and what was learned
tests/ 34 checks, and a diff against a PyTorch oracle given a model
serve/ 149 more for the server, incl. a differential vs upstream
examples/ chat.json for K3 and ChatML, the format a container carries
third_party/ stb_image.h, the single vendored header — see its README
docs/LEARNED.md is the one to read before contributing. It records what was measured, including the optimizations that were refuted — index-layout blocking, a 3-bit trunk, GPU offload, per-expert bit allocation — with the numbers that killed them.
Version 0.6.0, and the API is not frozen. Stated plainly, because finding these out for yourself is worse than reading them here:
- a container carries its chat format in
chat.json, and the converter can only fill it in for a model whose format has been transcribed from its reference encoder — K3 today. Neither Kimi release distributes a template, so for anything else the CLI says so and continues raw rather than guessing a format, which would produce plausible wrong answers instead of visibly wrong ones. Kimi-Linear is in that position now; - AVX-512 compiles and is dispatched from CPUID, and has still never
executed an instruction. This laptop is ARM and its x86 emulation is
Rosetta, which reports AVX2 and leaves the ZMM state disabled in XCR0;
the hosted x86 runner is an AMD EPYC 7763, which answers
avx512f/bw/dq/vl: no, so CI says AVX2 as well — on Linux and on Windows both. The workflow prints the runner's flags before every build, so the day a runner has them the SIMD backend matches the CPU baseline check becomes the confirmation without anyone arranging it; - Windows builds and runs, on one toolchain and one CPU. MinGW-w64
x86_64, cross-compiled, with
src/platform.hholding the six calls that are not POSIX: the positional read, the aligned allocation, the CPU count, the file size andFILE_FLAG_NO_BUFFERINGfor the cache-bypass open. MSVC is a different port and has not been attempted — the sources use GNU C. ARM64 Windows is not built. Neither is the page-cache bypass proven under load there: CI confirms Windows grants it on the runner's filesystem, which is not the same as measuring a hit rate against a container that does not fit in RAM; - the expert checksum is off unless you ask for it (
--verify), and the trunk has no checksum at all. The first is a decision — 5% of throughput on every token, against a container that is usually fine — and it means the default build of a rotted container still answers with whatever the damaged bytes decode to. Run--verifyonce after copying a container, to establish that it arrived intact.tools/verify_container.pydoes not stand in for that: it re-derives records against the source weights, so it wants torch and the original checkpoint on disk, and it answers whether the conversion was right rather than whether the copy still is. The second is not a decision: the trunk and the codebooks have nothing to check against in the format, and nothing has been built in its place; - every expert in a container is at the same bit width. The non-uniform per-expert allocation the format was designed around is not coming: it was measured on both models rather than built, and the importance it would allocate against does not vary — the value of the third bit spreads at most 1.15x between experts in a layer and 1.01x between layers, so the optimal allocator and a coin flip write the same container. The one signal that is not flat, routing frequency, buys disk footprint and almost no I/O, which is the resource that is actually scarce. docs/LEARNED.md §20 has the table and the one measurement that would revive it.
Apache 2.0 — see LICENSE. Copyright 2026 SQLite Cloud, Inc.