Changelog¶

v1.7.1 (Sep. 10, 2026)¶

New features and support¶

Add where() function.
Add set_printoptions() to set the number of terms and significant digits shown.
Add the alpha parameter to QAOA and RQAOA, which makes the cost the Conditional Value-at-Risk (CVaR) of the measured energies.
Add objective_format to FixstarsClient, AmplifyAEClient, and ToshibaSQBM2Client. The default Auto sends whichever of a dense matrix and a list of terms is smaller. Give Dense to keep the earlier behavior.
Add __repr__ and __str__ to AcceptableDegrees, VariableGenerator, split, and the key, value, and item views of Values and of the intermediate mapping. Each of them showed only its address before.
Support OQTOPUS Cloud with OqtopusClient.
Support Pyodide with a WebAssembly wheel that micropip installs and a browser runs. Only the formulation API and AmplifyAEClient work there, and each wheel fits one version of Pyodide.
Support the dense coupling mode of Amplify AE for a dense quadratic Poly, which needed a Matrix before.
Support file output for a model whose objective function is a Matrix.

Changes of behavior¶

Change objective to always return a Poly. A model now copies the Matrix that you give it, and it accepts the addition or the subtraction of a Poly.
Change a model to drop the variables that a Matrix only declares. variables lists fewer variables, and the problem that the client sends becomes smaller.
Change input_matrix and logical_matrix to raise a NotImplementedError. Use input_poly and logical_model_poly instead.
Change AmplifyAEClient to warn when you give a weight to a constraint and solve with the default Constraint solver. Amplify AE sets the weight itself, so the tutorials and the examples no longer give one.
Change a bound that narrows a Binary or an Ising variable to raise a ValueError. Amplify dropped such a bound before, so it solved another model than the one you described. Use equal_to() to fix a variable to one value.
Change a lower bound of positive infinity, or an upper bound of negative infinity, to raise a ValueError, because such a bound leaves no value. Amplify read it as “no bound” before, which widened the variable.
Change load_lp() and load_qplib() to give a constraint for a bound that narrows a Binary variable, such as x <= 0. Amplify dropped such a bound in silence before. Give the constraint enough weight, because a solver turns it into a penalty.
Change load_qplib() to read an integer variable whose bounds are 0 and 1 as a Binary variable, because QPLIB writes a binary variable in that form. Such a model held integer variables and no binary variable before.
Change save_qplib() to warn when a variable bound or a constraint bound reaches an absolute value of 1e+20. QPLIB reads that value as infinity, so loading the file back drops the bound, or fails when the sign of the bound leaves nothing that satisfies it.
Change save_lp() and save_qplib() to raise a ValueError for a coefficient or a constraint bound that is NaN or infinity. Amplify wrote nan or inf into the file before, and neither format gives those a spelling.
Change load_lp() and load_qplib() to raise a RuntimeError for a value that is not a number, such as nan. load_qplib() read such a value as zero in silence before, and load_lp() gave a model that no solver accepts.
Change an integer parameter to raise a TypeError for a value that loses precision, such as numpy.float64(3.0). It truncated such a value in silence before, and a NumPy integer, such as numpy.int64(3), still works.
Change set_num_threads() to limit the logical cores that Amplify uses, the parallel formulation included. Amplify caps a value that is larger than the number of the cores of the hardware.
Change set_num_threads() and set_seed() to take a named argument, num_threads and seed. A positional call still works.
Change amplify.Dim0 through amplify.Dim4 from classes to aliases of tuple, and add N for one axis. PolyArray[Dim2] still works, and only isinstance with one of these names stops.
Change the type stub to give the dimension of the result of an index, of a matrix product, and of an elementwise operation. A PolyArray that you build from a NumPy array or a list takes PolyArray[Dim], which no PolyArray[Dim2] accepts.
Change the D-Wave solver names to agree with the D-Wave API. A solver name no longer holds the system revision, so Advantage_system4.1 is now Advantage_system4.
Remove amplify.NECVA2Client and amplify.client.NECClient. NEC Vector Annealing (VA) Service 2.0 ended on April 1, 2025, so the client reaches no machine.
Raise a ValueError when you solve a model whose objective function holds no variable.

Performance¶

Improve performance of model formulation, including einsum(), one_hot(), and to_unconstrained_poly(). The TSP formulation is now up to 30 times faster than v1.0. See Formulation Benchmarks.
Improve memory usage of model formulation. The TSP formulation now needs approximately one-fourth to one-fifth of the memory of v1.0 for a problem with more than 10,000 variables.
Improve performance of to_intermediate_model() for a model with many integer or real variables, or with many higher-order terms.
Improve performance of load_qplib(). A QPLIB file of 230 thousand terms now loads about 1.2 times faster.
Update save_lp() and save_qplib() to raise a RuntimeError for invalid variable names or constraint labels, and emit warnings for ignored constraint weights or duplicate variable names.

Platform and dependencies¶

Update the D-Wave solver graphs for Advantage_system4, Advantage_system6, and Advantage2_system1.
Change supported macOS to Sonoma or later.
Update the required version of typing_extensions to >= 4.16.0, and of amplify-quantum to >= 1.3.0.

Documentation¶

Publish the documentation in Markdown for AI agents. Every page has a Markdown copy next to it, and each language also carries llms.txt and llms-full.txt.
Add an English version of Introduction to the latest version, which introduces the latest Amplify AE.
Add a description of every operator of Poly, PolyArray, and ConstraintList to the API reference and to help(), and remove the signatures of a PolyArray operation that no call could reach.
Add a description of the settings that every client carries, such as url, token, and parameters, and of each member of PenaltyFormulation.

Fixes¶

Fix: variables and get_variables() now list every variable of a Matrix objective whose first variable index is not zero.
Fix: load_lp() and save_lp() now support non-ASCII characters, such as Japanese, and a variable name of up to 255 characters, the indices included.
Fix: load_lp() now reads a file that has no objective section, and gives a model whose objective is 0. It raised an error before.
Fix: load_lp() now takes the type of a variable from the last variable type section that names it, as the LP format defines. A variable in both Binary and Generals became Binary before.
Fix: load_lp() now reads a variable name that starts with the spelling of a number, such as info or nano. Such a file failed to load before.
Fix: load_lp() now reads a bound that fixes a variable to one value, such as x = 3. Such a file failed to load before.
Fix: load_lp() now reads a file that has no End line, which the LP format allows. Such a file failed to load before.
Fix: save_lp() now raises a RuntimeError for a variable named inf, infinity, or nan. It wrote a file that no reader can load before.
Fix: save_lp() now keeps the exact value of every coefficient and bound. It wrote 10 significant digits before, so 0.1234567890123456 came back as 0.123456789.
Fix: save_lp() now writes -inf for a variable that has no lower bound. The LP default of a lower bound is 0, so x <= 3 alone gave the file another variable than the one you saved.
Fix: save_lp() now writes the bound of a Binary variable as 0 <= q <= 1. It wrote q free before, so the file gave a variable with no bound instead of a Binary one.
Fix: save_lp() now writes a constraint that holds a constant with the right sign. x + 5 <= 10 became x <= 15 in the file, which is another constraint than the one you gave.
Fix: save_lp() now writes every line within the 999 characters that the LP format allows, and it never breaks a bound over two lines. A file that held a long name failed to load.
Fix: load_qplib() now drops a constraint bound that reaches the infinity value of the file, as the QPLIB format defines. Such a bound became a real constraint before. A bound whose sign leaves nothing that satisfies the constraint now raises.
Fix: save_qplib() now keeps the exact value of every coefficient and bound. It wrote 6 significant digits before, so 0.1234567890123456 came back as 0.123457.
Fix: save_qplib() now writes a two-sided constraint that holds a constant with the right value on each side. 1 <= x + 5 <= 10 became 1 <= x <= 5 in the file, which is another constraint than the one you gave.
Fix: save_qplib() now writes the two indices of a quadratic term in the lower triangle, as the QPLIB format defines. Another reader could not load the file before.
Fix: save_qplib() now accepts a name that holds !, %, or # after its first character. It raised a RuntimeError for every such name before.
Fix: A model with an integer variable whose range exceeds 2147483647 now converts to the correct binary encoding. The solution of such a model did not solve the model that you gave before. A range that no encoding can build now raises an OverflowError.
Fix: to_intermediate_model() now gives a solution that is optimal for the input model when the quadratization method is IshikawaKZFD and the objective function holds Ising variables of degree three or higher.
Fix: to_intermediate_model() now raises a ValueError when the objective function and the constraints need different variable types. It returned a model of the forbidden variable type before.
Fix: An inequality constraint now gives a penalty function when Amplify cannot estimate the range of its left side, such as a left side that holds an unbounded variable. The penalty held NaN coefficients before, or it was 0, which removed the constraint from the model.
Fix: A constraint whose left side holds a squared term, such as a * a <= 3 with a in [-5, 5], now stays in the model. Amplify squared the two ends of the range only, so the constraint looked infeasible and its penalty became 0, or negative for an equality.
Fix: &=, |=, and ^= now update a PolyArray in place. The name became None before when the right side was a PolyArray, and it took a copy otherwise.
Fix: roll() now returns a copy of an array that holds no element. It ended the process before.
Fix: The display of a PolyArray no longer puts a space at the end of a wrapped line.
Fix: amplify no longer gives the names that its own implementation imports, such as amplify.sys and amplify.Never. from amplify import * gave them away, and none of them was ever documented.
Fix: The type stub of sum() and sum() now accepts axis for a zero-dimensional PolyArray. The call always worked, and only the stub rejected it.
Fix: The type stub of clamp() now names the second parameter bounds when you give axis for a PolyArray. A call with that keyword failed the type check although it always worked.
Fix: The type stub of the type parameter of HitachiClient no longer accepts None. The parameter never took None, and such an assignment raised a TypeError.
Fix: repr() of a Result that holds no solution now names the class, as it does for a result that holds one. It gave a text that began with { and ended with an unmatched ) before.

v1.6.0 (May. 21, 2026)¶

Support quantum computers via the Amplify Quantum extension.
Support Ubuntu 26.04.
Support D-Wave Advantage2_system1.
Fix: Correct response data not being written to file with write_response_data enabled in specific cases.

v1.5.0 (Apr. 05, 2026)¶

Overhaul the docstrings.
Support dense_objective endpoint of Amplify AE v1.1.
Support Gurobi Optimizer 12.0.3 on GurobiClient.
Change default degree reduction algorithm to Substitute from IshikawaKZFD.

v1.4.0 (Sep. 19, 2025)¶

Support Amplify Annealing Engine v1.0 on AmplifyAEClient.
Support D-Wave Advantage2_system1.6.
Deprecate D-Wave Advantage2_prototype2.6.
Rename AmplifyClient to BaseClient.
Support D-Wave solver identification updates and update the dependent dwave-system package to v1.33.0.
Update type hints for PolyArray and sum().
Fix: Substitute method for high-order polynomials now correctly reduces the degree to 2.
Fix: einsum() now correctly handles integer NumPy arrays.
Fix: FujitsuDA4Client now correctly handles inequality constraints without constants.

v1.3.0 (Nov. 01, 2024)¶

Improve support for type annotations.
Improve accuracy of upper and lower bounds estimates in constraint construction.
Change advanced parameters of solve(), parallel_solve(), equal_to(), less_equal(), greater_equal() and clamp() functions to keyword-only arguments for usability.
Support Hitachi CMOS annealing on HitachiClient.
Support D-Wave Advantage2_prototype2.5.
Support conversion from Variable to Poly.
Support input/output of models with LP and QPLIB format files.
Variable can be used as a mapping key for substitute(), substitute() and is_satisfied().
Support conversion from real variables to binary variables.
Add compress_qpu_problem_data attribute to DWaveSamplerClient, LeapHybridSamplerClient and LeapHybridCQMSamplerClient.
Add support for Python 3.13, and Python 3.8 is no longer supported.

v1.2.0 (July 23, 2024)¶

Support TOSHIBA SQBM+ v2.0.3 on ToshibaSQBM2Client.

v1.1.0 (May 13, 2024)¶

Support NEC Vector Annealing Service 2.0 on amplify.NECVA2Client.
Improve the efficiency of the Substitute quadratization method and significantly reduce the number of auxiliary variables.
Add support for Linux ARM64.

v1.0.0 (Jan. 27, 2024)¶

Release version 1.0.