What is Bend2?
Last updated .
Bend2 is the in-development successor to Bend, Higher Order Company’s massively parallel programming language. There is no public code as of 2026-08-04, and a countdown posted by creator Victor Taelin points at 2026-08-12.
The claim
Bend2, per Taelin’s posts through mid-2026, combines three properties that rarely coexist in one language: ordinary-looking functional code that parallelizes automatically, a dependent type system strong enough to state and prove properties of that code, and program synthesis that generates implementations from sufficiently precise types. Each piece exists somewhere already. Automatic parallelism exists in array languages, dependent types in Lean and in Higher Order Company’s own Kind, synthesis in a decade of research prototypes. The bet is the composition: a type precise enough to pin down behavior turns generated code into something a checker can accept or reject, and a runtime that extracts parallelism on its own makes the accepted code fast without hand-scheduling. None of this is verifiable against public code yet. This page is a map of claims, each carrying a date.
Why interaction nets parallelize
Bend2’s performance model rests on the same substrate as Bend1’s: interaction nets. A program is a graph, and evaluation is a small fixed set of rewrite rules, each of which touches exactly two adjacent nodes. Two useful properties fall out of that locality. First, rewrites that share no nodes cannot conflict, so every independent redex in the graph can fire concurrently without locks. Second, the rules are confluent, meaning the final result does not depend on the order in which redexes fire, so a scheduler can hand work to any idle thread, or any idle GPU warp, without coordination. That is the entire trick behind running an unannotated functional program across thousands of cores. The price is the constant factor: every step is graph surgery with real memory traffic, and that cost is where the criticism of Bend1 concentrated at its launch.
The runtime lineage
Bend2 targets the fourth generation of a runtime line that Higher Order Company has been iterating on since 2022: HVM1, the original lazy evaluator for interaction combinators, then HVM2, the strict rewrite that compiled to C and CUDA and carried Bend at its May 2024 launch, then HVM3, and now HVM4, which is public and saw pushes into mid-2026. On 2026-07-18 Taelin declared the language done and reported that Bend2’s CUDA runtime had been implemented overnight by a coding model (Anthropic’s Fable) from the reference Metal runtime, reusing most of the existing code, and ran faster than Metal on RTX, a baseline he put at roughly 10x parallel C for most programs. The engineering read is unglamorous: porting a runtime against a working reference is a well-specified task, because the reference doubles as an executable spec and a differential test oracle. The interesting decision was building the Metal version first.
What exists in public
Bend2’s public footprint, as of 2026-08-04, is an empty repository. VictorTaelin/Bend2 on
GitHub carries the description “Bend2 - WIP”, zero bytes of code, and no branches. There is no
documentation, no binary, no package on any registry, and no license text. The adjacent public
artifacts are Bend, which still receives maintenance,
and HVM4. The clearest picture of what remained before
launch is Taelin’s own
2026-03-05 list: HVM4’s AOT compiler,
HVM4’s GPU runtime, and Bend2-SupGen integration, with the note that once those were ready
“all the rest is trivial”. SupGen is the synthesis component’s name in his posts.
Where this goes next
Bend2’s release is expected around 2026-08-12. The arithmetic behind that date and its failure modes are covered in When does Bend2 release?, the comparison against what Bend1 actually shipped is in Bend2 vs Bend1, and the front page holds the live yes/no answer. This page gets rewritten against the compiler the week it ships.