---
type: "post"
id: "runanywhere-releases-trnrt-0-1-0-its-first-818c75"
slug: "runanywhere-releases-trnrt-0-1-0-its-first-818c75"
title: "RunAnywhere releases TrnRT 0.1.0, its first public release for AWS Trainium"
description: "RunAnywhere’s TrnRT 0.1.0 changelog, dated November 8, 2026, calls this the serving engine’s first public release."
retrieval_nugget: "RunAnywhere’s TrnRT 0.1.0 changelog, dated November 8, 2026, calls this the serving engine’s first public release. It describes a GLM-5.3-Flash deployment on one AWS Trainium2 instance, an OpenAI-compatible API and newly added speculative-decoding modes. For builders, the release makes a concrete Trainium serving stack inspectable and testable."
published_at: "2026-11-17"
updated_at: "2026-11-17"
record_date: "2026-11-17"
date_kind: "published_at"
topics: ["inference","ai-infrastructure","compute"]
entities: ["RunAnywhere","TrnRT","GLM-5.3-Flash","AWS Trainium2","Amazon Web Services"]
related_patterns: []
novelty_label: "new_fact"
what_changed: null
prior_refs: []
source_url: "https://github.com/RunanywhereAI/trnrt/blob/main/CHANGELOG.md"
source_urls: ["https://github.com/RunanywhereAI/trnrt/blob/main/CHANGELOG.md"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:runanywhere-releases-trnrt-0-1-0-its-first-818c75"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/runanywhere-releases-trnrt-0-1-0-its-first-818c75/","markdown":"https://newruntime.com/posts/runanywhere-releases-trnrt-0-1-0-its-first-818c75.md","json":"https://newruntime.com/posts/runanywhere-releases-trnrt-0-1-0-its-first-818c75.json"}
markdown_url: "https://newruntime.com/posts/runanywhere-releases-trnrt-0-1-0-its-first-818c75.md"
---

# RunAnywhere releases TrnRT 0.1.0, its first public release for AWS Trainium

## Retrieval answer

RunAnywhere’s TrnRT 0.1.0 changelog, dated November 8, 2026, calls this the serving engine’s first public release. It describes a GLM-5.3-Flash deployment on one AWS Trainium2 instance, an OpenAI-compatible API and newly added speculative-decoding modes. For builders, the release makes a concrete Trainium serving stack inspectable and testable.

## Summary

RunAnywhere’s TrnRT 0.1.0 changelog, dated November 8, 2026, calls this the serving engine’s first public release. It describes a GLM-5.3-Flash deployment on one AWS Trainium2 instance, an OpenAI-compatible API and newly added speculative-decoding modes. For builders, the release makes a concrete Trainium serving stack inspectable and testable.

## What happened

RunAnywhere’s changelog identifies TrnRT 0.1.0 as the first public release and dates it November 8, 2026. Its headline configuration serves GLM-5.3-Flash on one trn2.48xlarge. It uses two engines, each with 32 tensor-parallel ranks, behind a session-affine router. The server exposes an OpenAI-compatible API, supports a one-million-token context, and includes a prefix cache with a host-memory tier and NKI decode megakernels. The changelog explicitly says these details describe the configuration, not a measurement; it points readers to `docs/RESULTS.md` for measured numbers and their context, and `docs/DEPLOY.md` for deployment.

The release adds multi-token-prediction speculative decoding using GLM-5.3-Flash’s own MTP layer. TrnRT drafts two tokens in a step, then verifies them with the target model in the next step. The scheduler runs that work before the previous step is read back. Users can select the `mtp-always`, `mtp-batch8`, `mtp-batch12` or `mtp-batch24` presets. A dynamic mode speculates only while few requests are decoding. The release also adds `trnrt:spec_decode_*` metrics, giving operators a way to observe the feature. The changelog directs users to `docs/CONFIGURATION.md` for presets and `docs/AGENTX.md` for benchmark details.

One hardware-specific detail matters when interpreting the published MTP results: `TRNRT_DMA_HW` enables the kernels’ static DMA helpers in hardware DGE. The MTP presets set it, as did the builds behind those published numbers; other configurations leave it unset. The changelog also adds a synthetic benchmark mode, `--spec-synth-al`, to reproduce InferenceX’s synthetic acceptance, but says it is off by default. That distinction separates a reproducible synthetic acceptance setup from the ordinary default behavior.

The public release includes a README, documentation, contribution and security policies, issue and pull-request templates, and a release workflow. The changelog lists CI checks including pre-commit linting and formatting with Ruff, mypy, a CPU test suite, package build checks, Dockerfile lint and CodeQL. It describes a review and promotion process: pull requests target `development`, require maintainer approval and resolved review conversations, while `main` takes release promotions. A version-bumping promotion to `main` triggers a version tag and publication once per version. The changelog also records fixes for router event-loop stalls on large requests, aborted engine work after a non-streaming client disconnects, and invalid JSON request bodies returning 400 rather than 500.

## What the archive adds

New Runtime covered RunAnywhere’s earlier TrnRT release on November 5, 2026, describing its Trainium2 serving engine for GLM-5.3-Flash and its NKI layer megakernels ([earlier note](https://newruntime.com/posts/runanywhere-open-sources-trnrt-for-serving-glm-5-39e01e/)). This is a follow-on release, not a restatement: the new fact is the first public 0.1.0 release, including documented speculative-decoding modes and the associated configuration details. Earlier New Runtime coverage also examined making Hugging Face models easier to run on Trainium through TorchNeuron ([October 22 note](https://newruntime.com/posts/torchneuron-running-hf-models-on-aws-trainium-is-787908/)).

## Trend signal

This release does not yet fit a pattern New Runtime tracks. It adds a specific serving implementation and configuration for one model on Trainium2, but the changelog alone does not establish a broader shift in agent architecture, interoperability, or the economics of verified completed work. Treat it as a concrete release to evaluate on its stated hardware and workload, rather than evidence for a general trend.

## Why now

The changelog dates the release to November 8, 2026, and labels it the first public release. That timing marks a transition from the earlier described TrnRT project to a versioned release with public documentation and contribution processes. The changelog does not explain why the team chose this date or identify a deadline or market event behind the publication.

## Why it matters

For teams already evaluating Trainium2, this release provides a specific software path to test rather than only a hardware target: model, instance type, parallelism layout, router, API surface and cache design are named together. The MTP options also expose operational choices that can be measured against a team’s own request mix. A preset, dynamic speculation, and the hardware DMA setting are not interchangeable descriptions; published results were produced with the MTP presets and `TRNRT_DMA_HW` enabled.

The value is therefore practical but bounded. Teams can inspect the release, reproduce its deployment path, and compare the documented measurements with their own workload, while retaining an OpenAI-compatible client interface. They should not infer speed, cost advantage, production readiness or parity with another serving stack from the configuration list alone. The changelog’s explicit distinction between setup and measured results is important: deployment architecture is a starting point for evaluation, not a performance claim.

## Limits and uncertainty

The changelog excerpt gives no numerical latency, throughput, cost, or acceptance-rate results; it directs readers to `docs/RESULTS.md` for measurements and context. It does not establish how the described configuration performs across workloads, under different concurrency levels, or against other serving engines. It also does not state service-level guarantees, production support terms, or the operational limits of the one-million-token context. The deployment documentation and benchmark results require separate review before a team can decide whether the setup meets its needs.

## Skill to build now

If your team has Trainium2 access and serves GLM-5.3-Flash, reproduce the documented baseline before tuning: record the exact instance, two-engine/32-rank layout, router behavior, cache settings, and whether `TRNRT_DMA_HW` is set. Then compare an MTP preset with the dynamic mode under your actual request mix, using the `trnrt:spec_decode_*` metrics. Keep synthetic acceptance runs separate from production-like measurements, since the changelog says the synthetic mode is off by default.

## Reader takeaway

TrnRT 0.1.0 turns RunAnywhere’s Trainium2 serving work into a public, versioned implementation with an OpenAI-compatible API and configurable MTP speculative decoding. The useful next step is not to assume its headline configuration is faster, but to reproduce it and examine the published measurements against your own workload and hardware setup.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Inference](https://newruntime.com/topics/inference.md): Explore the inference topic hub.
- topic hub: [Ai Infrastructure](https://newruntime.com/topics/ai-infrastructure.md): Explore the ai-infrastructure topic hub.
- topic hub: [Compute](https://newruntime.com/topics/compute.md): Explore the compute topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
