Tsfm.Forecasting.Timestral.Weights.28M 0.2.0

Tsfm.Forecasting

Probabilistic forecasting with native .NET inference for time-series foundation models, on CPU or GPU via TorchSharp. Currently Kronos, TimesFM 3.0, t0 (t0-alpha, t0-beta) and Timestral (5M, 28M, 123M, 209M).

Targets .NET 10

NuGet packages:

Package	Licence
Tsfm.Forecasting	MIT	safetensors reader, checkpoint abstraction
Tsfm.Forecasting.Kronos	MIT	tokenizer, model, forecaster
Tsfm.Forecasting.Kronos.Weights.Small	MIT	Kronos-small (24.7M) + Tokenizer-base
Tsfm.Forecasting.Kronos.Weights.Mini	MIT	Kronos-mini (4.1M) + Tokenizer-2k
Tsfm.Forecasting.TimesFm	Apache-2.0	TimesFM 3.0 — no weights
Tsfm.Forecasting.T0	Apache-2.0	t0-alpha, t0-beta — open weights, fetched separately
Tsfm.Forecasting.Timestral	Apache-2.0	Timestral (Mistral AI) — code only
Tsfm.Forecasting.Timestral.Weights.5M	Mistral Modified MIT — revenue-restricted	Timestral-5M (5.2M)
Tsfm.Forecasting.Timestral.Weights.28M	Mistral Modified MIT — revenue-restricted	Timestral-28M (27.8M)

Refer to respective package for license and conditions of use (if any).

Why this exists

None of these models is a text LLM, so existing .NET inference runtimes cannot load them. Kronos pairs a Binary Spherical Quantization autoencoder with a decoder that has a hierarchical two-subtoken embedding, a dependency-aware cross-attention layer and two conditional heads. GGUF has no representation for any of that, which rules out llama.cpp-based runtimes; the alternatives that do reach Apple's GPU load GGUF only. The architectures had to be ported.

Kronos — batteries included

dotnet add package Tsfm.Forecasting.Kronos
dotnet add package Tsfm.Forecasting.Kronos.Weights.Small   # or .Mini
dotnet add package TorchSharp-cpu                          # or a CUDA backend

using Tsfm.Forecasting.Kronos;
using Tsfm.Forecasting.Kronos.Weights;
using static TorchSharp.torch;

using var forecaster = KronosForecaster.Load(KronosSmall.Instance, new Device("mps"));

var rows = KronosForecaster.OutputCount(barCount, contextBars: 384);
Span<float> lean = new float[rows];
Span<int>   upCount = new int[rows];

forecaster.Infer(
    ohlcva,          // row-major [L x 6]: open, high, low, close, volume, amount
    barTimeMs,       // [L] Unix ms, one per bar
    lean, upCount, dispersion: default,
    contextBars: 384, horizon: 1, rollouts: 30,
    greedy: false, temperature: 1f, topP: 1f);

Supplying fewer than six channels

The model always reads six channels and has no mask, so an absent one is filled, not ignored — whatever you supply is read as data. There is a single entry point rather than ohlc/ohlcv overloads, because choosing the filler is a modelling decision and belongs at the call site. These are the fills the reference implementation uses:

Note the mean of all four prices, not the close. The reference rejects NaN in any of the six rather than treating it as absent, so fill explicitly.

Checkpoints are embedded as assembly resources, so nothing resolves through configuration or the filesystem — which is what lets this sit behind a consumer forbidden from reading its environment. Buffers are caller-supplied and sized from OutputCount; a mismatch throws with the expected count rather than silently misaligning every row.

TimesFM — bring your own weights

The weights are not open source and are not distributed here. TimesFM 3.0 checkpoints are published under the TimesFM Non-Commercial License v1.0: research and evaluation only, with revenue-generating activity and production deployment expressly forbidden. This package's Apache-2.0 licence grants no rights in them whatsoever. Commercial use needs terms from Google.

dotnet add package Tsfm.Forecasting.TimesFm
dotnet add package TorchSharp-cpu
./scripts/fetch-timesfm-checkpoint.sh    # ~1.2 GB — read the licence first

using Tsfm.Forecasting.TimesFm;

var forecaster = TimesFmForecaster.Load("checkpoints/timesfm-3.0-pytorch", new Device("mps"));
double[,] q = forecaster.Forecast(ohlcva, horizon: 4);   // [step, quantile]

One forward pass yields 64 steps at all nine quantiles, so a full predictive interval costs no more than a point forecast. Every OHLCVA channel is supplied as a variate and marked a target: preprocessing masks future covariate slots only for target variates, so a channel left non-target would be handed its own future values.

t0 — open weights, fetched

t0 from The Forecasting Company is a decoder-style patch transformer alternating causal time attention with variate (group) attention, with native known-future covariates. Both checkpoints are Apache-2.0, but the package carries no weights: install them yourself, as below.

dotnet add package Tsfm.Forecasting.T0
dotnet add package TorchSharp-cpu

Installing the weights

T0Forecaster.Load(dir, device) needs a directory holding two files from the Hugging Face repository, unrenamed:

<dir>/config.json          # architecture and scaler convention — required, not optional
<dir>/model.safetensors    # 407 MB (alpha) or 1.0 GB (beta)

From a clone of this repository, the script fetches both checkpoints at the revisions the port was verified against, and checks each payload's SHA-256:

./scripts/fetch-t0-checkpoint.sh            # both, into checkpoints/t0-alpha and checkpoints/t0-beta
./scripts/fetch-t0-checkpoint.sh t0-beta    # or one

From a project that only references the package, download them directly — pinning the revision, since a refreshed checkpoint would silently change every forecast:

fetch() {   # name, revision, sha256
  mkdir -p "models/$1"
  for f in config.json model.safetensors; do
    curl -fL -o "models/$1/$f" "https://huggingface.co/theforecastingcompany/$1/resolve/$2/$f"
  done
  echo "$3  models/$1/model.safetensors" | shasum -a 256 -c
}
fetch t0-alpha 741927847fe02367624749bb9290b2f09f128e7d 16c030d3fd70f06dc4238e9a8356e9b5a631d07f80f1bc76ba539991aed5897f
fetch t0-beta  941e522547ec63f0f2aad57089fa0f5b8498bdf8 a0fd8abd51275dd30afd21f4892c763ebd45e2cb1811eff26eed43f89a543f7d

hf download theforecastingcompany/t0-beta config.json model.safetensors --revision <rev> --local-dir models/t0-beta does the same through the Hugging Face CLI. Deploy the directory beside your application like any other data file; nothing is read from the environment or a cache, and a missing file fails at Load with these instructions in the message.

Use

using Tsfm.Forecasting.T0;

var t0 = T0Forecaster.Load("checkpoints/t0-beta", new Device("mps"));

// Same shape as TimesFM: variate-major series, known-future flags, target index.
double[,] q = t0.ForecastJoint(series, variates: 3, contextBars: 512, horizon: 24,
                               knownFuture: [false, true, true]);   // [step, quantile]
double[,] r = t0.Forecast(ohlcva, horizon: 4);                      // forward returns

Both forecasters implement IQuantileForecaster, so one call site drives either; Quantiles names the columns, whose count differs per checkpoint. samples/GroceryCovariates runs on either (-- t0-beta). Both also batch; see Batching.

What differs from TimesFM:

• Any context length. The context is left-padded onto the 32-step patch grid and the padding masked, so nothing need fill whole patches. No ceiling is imposed; the reference evaluates at 8192.
• Up to 1024 steps per pass. Every horizon patch is decoded at once. The reference continues past 1024 by an autoregressive rollout over quantile paths, which is not ported.
• Information flows from covariates to targets, never back. In variate attention a target reads every series in its entry, but a known-future series reads only the known-future series (itself included). A covariate's representation therefore never depends on the target's history, and cannot smuggle it forward. Known-future series are also scaled by global rather than causal statistics.
• Each checkpoint names its own scaler convention. t0-beta's config.json carries scaler_eps and scaler_eps_mode; alpha's predates them, so their absence means alpha's values. Running beta under alpha's convention loads cleanly and silently degrades the forecast — which is why they are read rather than assumed.

Timestral — 5M and 28M batteries included

The weights are not open source. Mistral AI publishes them under a Modified MIT licence whose condition 2 withholds every right from anyone whose company — or employer — had consolidated monthly revenue above USD 20 million in the preceding month. That binds each consumer, not only whoever publishes a package. The weights packages carry the licence verbatim (LICENSE.txt, the package licence, and Timestral5M.License at run time) and require its acceptance on install. The code here is Apache-2.0 and grants nothing in the weights.

Timestral is a causal patch transformer with time and variate attention, trained with Decoupled Horizon Forecasting: forecast-query tokens read the encoded context, and a monotone head emits 21 quantiles per step. Context up to 32,768 steps; up to 1,024 forecast in one pass.

dotnet add package Tsfm.Forecasting.Timestral.Weights.5M     # or .28M; brings the code package
dotnet add package TorchSharp-cpu

using Tsfm.Forecasting.Timestral;
using Tsfm.Forecasting.Timestral.Weights;

var ts = TimestralForecaster.Load(Timestral28M.Instance, new Device("cpu"));
double[,] q = ts.ForecastJoint(series, variates: 3, contextBars: 2048, horizon: 96,
                               knownFuture: [false, true, true]);    // [step, 21 quantiles]

Timings are 4 CPU cores with the defaults below. The port is faster than the reference mainly because it folds the mirrored input of symmetric averaging into the same forward pass, as extra groups, rather than making a second call. The larger two load from a directory: TimestralForecaster.Load("checkpoints/Timestral-209M", device), or any IModelWeights. The fetch script pins revisions and checks SHA-256; the weights packages check the same hashes at build time and refuse a payload that differs.

The forecaster applies what the reference forecaster applies by default, so published numbers are reproduced:

• Symmetric averaging (SymmetricAveraging, on): forecast the negated input too, with every series negated, and average the reflection in.
• Non-negative clipping (ClipNonNegative, on): a series whose observed context is never negative is clipped at zero. Harmless on prices; never triggers on returns.
• Advancing refine frame: horizon patch j is denormalised with the statistics of the context plus the block's own medians for patches before it, the frame it was trained in.
• Quantile cap: forecasts are held within 10⁴ context scales of the observed range.

Known-future series are scaled by whole-row statistics, and a series observed only at 0 and 1 is treated as an indicator. A context beyond 32,768 steps is refused rather than silently truncated, as Kronos's is. Not ported: past-only covariates as a separate role (a series not marked known is a joint target), categorical target encoding, and the block-by-block continuation past 1,024 steps.

Batching

Each forecaster has a batch form of its single call. Entries share a horizon and nothing else — they never attend to one another — so a batched result equals the single call's to float32 rounding. Not bit for bit in general: whether a matmul over more rows reproduces one over fewer is up to the BLAS, and macOS's does not. For Kronos, whose sampling picks tokens by comparing a uniform against a CDF, a last-bit difference can in principle flip a token, so batch invariance there is exact only where the backend is. What batching buys is fewer, larger forward passes: one per anchor is the dominant cost of a backtest.

The t0, Timestral and TimesFM batches are on IQuantileForecaster, take JointSeries entries (or OHLCVA windows) as ReadOnlyMemory, and write [entry][step][quantile] into a span sized by BatchOutputLength. Anchors along one history are slices of one array, so building a backtest batch copies nothing:

var windows = anchors.Select(a => history.AsMemory((a - context + 1) * 6, context * 6)).ToArray();
var returns = new double[forecaster.BatchOutputLength(windows.Length, horizon)];
forecaster.ForecastBatch(windows, horizon, returns);

Batching and rollouts. The batch axis holds independent series; rollouts are samples within one series. Kronos flattens both into one tensor batch of series × rollouts rows, and a RolloutAggregator reduces only across the rollouts of one series, at one step and channel — never across series. So the two compose without interaction, except in cost: memory scales with the product, which is why Kronos's batch counts windows and leaves rows to be batch * rollouts. Sampling is seeded per window from its anchor bar, so results do not depend on how windows were grouped into passes. By default two instruments anchored on the same bar therefore draw the same uniforms — common random numbers, which make cross-sectional comparisons less noisy but leave their errors correlated. That is a host policy, KronosForecaster.Seed: SamplingSeeds.PerSeries gives each KronosSeries.SeriesKey its own stream, and SamplingSeeds.Salted derives a fresh, reproducible run from either. TimesFM, t0 and Timestral emit quantiles directly, with no rollouts and nothing to aggregate. Timestral's symmetric averaging is the nearest thing: the input and its negation travel as two groups of one pass, and are combined per series afterwards. t0's upstream continuation past 1024 steps (not ported) would be the same pattern again: each requested quantile becomes a path, paths are folded into the batch as extra groups, and they are reduced per series to quantiles.

Choose your own backend

Neither model package propagates a native runtime, because the right one is platform- and accelerator-specific. TorchSharp-cpu resolves it per RID; see TorchSharp's download guidance for CUDA.

Things that will bite you

DisposeScope is mandatory. TorchSharp has no refcounting. Every forward pass these libraries perform is scoped; if you write your own, wrap it in using var _ = torch.NewDisposeScope(). Without it the Metal allocator thrashes and you will measure roughly a 10x slowdown that looks like a backend problem and is not.

Cached tensors need DetachFromDisposeScope(), not MoveToOuterDisposeScope(). The latter hands ownership to the caller's scope, which frees it on exit — so a cache is released after its first use.

Metal is present in the "cpu" backend. On Apple Silicon "cpu" means not CUDA. Probe by placing a tensor rather than trusting the name.

macOS needs Homebrew's libomp. libtorch_cpu.dylib links /opt/homebrew/opt/libomp/lib/libomp.dylib by absolute path, so the copy shipped inside the NuGet package is never used. Without it, loading fails with a message claiming the backend reference is missing — which it is not.

No model has a key-value cache. Attention is recomputed over the whole context at every decode step, matching the references. A cache would change the arithmetic and forfeit checkable parity. For Kronos this costs horizon full passes per rollout, so long horizons are expensive; TimesFM and t0 emit their whole horizon in one pass and are unaffected.

Parity

Both ports are verified stage by stage against their references, on CPU and Metal, rather than judged on final outputs — agreement at the end cannot distinguish a wrong attention scale from a misplaced norm, since both merely shift the result.

Kronos. The tokenizer's embedding and first norm are bit-exact; relative error thereafter is 1e-7 to 1e-6, ordinary float32 accumulation. It is not bit-identical and does not claim to be: cross-implementation token disagreement extrapolates to near 1 in 15,000. Sampling differs deliberately — per-bar uniforms from a SplitMix64 stream seeded by the bar's own timestamp, selected by inverse CDF, so a draw does not depend on how bars were grouped into batches. The reference seeds one stream per batch and is not batch-invariant. Distributionally identical; the stream is not.

TimesFM. Relative error is ~1e-7 through preprocessing and the layer stack, ~2e-6 at the logits. Two faults the harness caught, neither of which throws and both of which produce entirely plausible forecasts: scaled_dot_product_attention ignores is_casual once attn_mask is supplied, so supplying a patch mask silently disabled causal masking and let every position attend to its own future; and the reference leaves rescale_logits false, making its logits QK^T * sqrt(d) rather than the conventional QK^T / sqrt(d).

t0. Every stage — scaling, attention masks, patch embedding, each layer, the quantile head, the rescale — agrees with tfc-t0 0.6.0 to float32 rounding: masks and scaling are bit-exact, layers within 6e-7 relative, forecasts within 5e-7. Each stage is fed the reference's previous output, so a fault cannot hide behind an earlier one. The end-to-end cases include the quickstart notebook's own examples: on the electricity-price series, t0-beta's median misses the truth by 6.522 without covariates and 2.067 with them (notebook: 6.528 without, recorded on different hardware), and the port reproduces the reference to the same digits. Three details would each have produced plausible forecasts if missed: rotary pairs are interleaved where TimesFM's are split in halves; the QK norm precedes the rotation, where TimesFM's follows it; and a padding patch has no key to attend to, so attention must return zero there, not NaN, or every row reading it is poisoned.

Timestral. All four sizes agree with mistralai/timestral 1.0.0 stage by stage: scaling, masks and positions bit-exact, every other stage within 1.5e-6 relative, and forecasts within 9e-7 — with the forecaster's defaults and with them off, and including the covariate path, the refine frame and the electricity-price and M4 examples shared with t0. The reference builds a decoder token for every patch of the window and discards those before the horizon; the port builds only the horizon's, which is exact because the discarded ones cannot reach the rest. Details that would each have forecast plausibly if missed: the fused QKV projection is laid out (qkv, head_dim, heads), not head-major; RoPE turns only the first half of each head, in the split-half layout; SwiGLU's SiLU falls on the second half, where t0's is on the first; and statistics are float64 Welford, which MPS cannot do, so they run on the CPU.

./scripts/fetch-checkpoints.sh                       # Kronos
dotnet test tests/Tsfm.Forecasting.Kronos.Tests -c Release

./scripts/fetch-timesfm-checkpoint.sh                # TimesFM, non-commercial
./scripts/fetch-timesfm-reference.sh
python3 -m venv .venv && ./.venv/bin/pip install torch safetensors numpy huggingface_hub
./.venv/bin/python reference/dump_timesfm_parity.py
dotnet run --project tests/Tsfm.Forecasting.TimesFm.Parity -c Release

./scripts/fetch-t0-checkpoint.sh                     # t0, Apache-2.0
./.venv/bin/pip install "tfc-t0[torch]==0.6.0" pandas pyarrow
./.venv/bin/python reference/dump_t0_parity.py checkpoints/t0-alpha
./.venv/bin/python reference/dump_t0_parity.py checkpoints/t0-beta
dotnet run --project tests/Tsfm.Forecasting.T0.Parity -c Release -- t0-alpha t0-beta

./scripts/fetch-timestral-checkpoint.sh all          # Timestral, revenue-restricted weights
./.venv/bin/pip install "git+https://github.com/mistralai/timestral.git@a534bdbd244d64f3d9d743f46c7bf6fc330ed4ff"
for m in 5M 28M 123M 209M; do ./.venv/bin/python reference/dump_timestral_parity.py checkpoints/Timestral-$m; done
dotnet run --project tests/Tsfm.Forecasting.Timestral.Parity -c Release -- Timestral-5M Timestral-28M Timestral-123M Timestral-209M

Building the packages

./scripts/pack.sh            # or: ./scripts/pack.sh <output-dir>

Fetches the checkpoints the weights packages embed, packs every package at the version in Directory.Build.props into artifacts/, verifies licences and payloads, and prints the order to upload in: each package depends on the exact version of its parent, so parents must be on nuget.org, and indexed, before their dependants. Nothing is pushed.

Attribution

Kronos. Model architecture, BSQ tokenizer and reference implementation by ShiYu (MIT). Published checkpoints are by NeoQuasar (MIT), redistributed unmodified in the Tsfm.Forecasting.Kronos.Weights.* packages, which record the revision each was built from.

TimesFM. Derived from google-research/timesfm, Copyright 2026 Google LLC, Apache-2.0. No Google source files and no weights are redistributed here; both are fetched at pinned revisions by the scripts above.

t0. Derived from tfc-t0, Copyright 2026 The Forecasting Company, Apache-2.0, which in turn builds on Datadog's Toto and Amazon's Chronos-2, both Apache-2.0. No weights or upstream sources are redistributed; checkpoints are fetched at pinned revisions.

Timestral. Derived from mistralai/timestral, Copyright Mistral AI, Apache-2.0. The 5M and 28M weights are redistributed unmodified under Mistral AI's Modified MIT licence, which is reproduced in full in the weights packages; the 123M and 209M weights are not redistributed.

This project is an independent port and is not affiliated with or endorsed by any of them. See NOTICE for the full attribution and the statement of changes.