---
title: "AI engineering, October 20, 2026"
description: "The Forward Pass daily issue, October 20, 2026."
canonical: https://theforwardpass.net/archive/daily/2026-10-20
updated: 2026-10-20
---

# AI engineering, October 20, 2026

The Forward Pass daily issue, October 20, 2026. Source: https://theforwardpass.net/archive/daily/2026-10-20

> This issue is researched and written by AI models, and every fact is checked against its cited source. No human edits it before it is sent.

_MiniMax cuts agentic costs, Inception ships Mercury 3, AWS brings PyTorch to Trainium, Unsloth runs 552B DeepSeek locally_

Inception's Mercury 3 reaches 1,068 tokens per second on NVIDIA GPUs. It's a diffusion-based language model.

Meanwhile, MiniMax prices M3.1 Flash at $0.60 per million output tokens. Deepgram reports about 60% more audio processed per dollar on Trainium than on its GPU baselines, as AWS introduces TorchNeuron.

If you read one thing today: TorchNeuron. See how to move existing PyTorch models onto Trainium using standard device selection and torch.compile.

## MiniMax ships M3.1 Flash with 70% lower agentic usage costs

**Top News** · 423 HN points

M3.1 Flash is MiniMax's hosted model for coding, computer use, and multi-step workflows, with output priced at $0.60 per million tokens. You can access it through the MiniMax API, M Plan, and MiniMax Code.

Here's what you get:
- Input costs $0.15 per million tokens, while cached input costs $0.015 per million tokens.
- Everyday output speeds reach up to 200+ tokens per second, with median speeds of 170+ tokens per second.
- Complex software task performance reaches 73.7% on DeepSWE v1.1.
- Computer-use results include 61.7 on OSWorld V2, alongside 29.6 on Agents' Last Exam.

One catch: Model weights are not available yet.

Try it: Use the MiniMax API at https://platform.minimax.io, or download MiniMax Code for desktop at https://agent.minimax.io/download.

Sources: [minimax.io](https://www.minimax.io/blog/minimax-m31f)

## Inception says Mercury 3 matches GPT-6 Luna while running 4-8x faster

**Top News** · 234 HN points

Mercury 3 is Inception's flagship diffusion-based LLM. Inception says it matches or beats GPT-6 Luna and Claude Haiku 5.5 while running 4-8x faster.

Here's what changed:
- Generation reaches 1,068 tokens per second on NVIDIA GPUs.
- Launch pricing is 50% below the standard rates of $0.20 per million input tokens and $0.70 per million output tokens.
- Cached input tokens cost 90% less than the input price.
- The API is OpenAI API compatible, and the model supports a 1-million-token context window, up from Mercury 2.5's 260,000 tokens.

One catch: The 50% Mind2Web completion result comes from Inception's own evaluation of 50 live navigation tasks.

Try it: Use the Inception API Platform with 100 million free API tokens, or access Mercury 3 through OpenRouter or Baseten. Also available via Vercel

Sources: [inceptionlabs.ai](https://www.inceptionlabs.ai/blog/introducing-mercury-3) · [inceptionlabs.ai](https://inceptionlabs.ai/blog/introducing-mercury-3)

## TorchNeuron brings PyTorch to Trainium, Deepgram reports 60% more audio per dollar

**Top News** · Reported · pytorch.org

If you want to move an existing PyTorch model onto Trainium, TorchNeuron gives you a native backend. Deepgram reports about 60% more input audio seconds per dollar on Trainium than on its existing GPU baselines.

Here's what it supports:
- Public beta is scheduled for October 20, 2026, for researchers and developers, with packages from the AWS Neuron pip repository.
- Trainium2 and Trainium3 support eager execution, torch.compile, and custom kernels through NKI.
- Distributed workloads can use DDP, FSDP2, tensor parallelism, and pipeline parallelism.

One catch: The first release excludes dynamic and symbolic shapes and sparse tensors, and some operators lack native lowerings.

Try it: Install `pip install torch-neuronx` from the AWS Neuron pip repository. Move your model with `model = MyModel().to('neuron')`, then compile with `torch.compile(model, backend="neuron")`.

Sources: [pytorch.org](https://pytorch.org/blog/introducing-torchneuron-native-pytorch-backend-for-aws-trainium/)

## Signals

1. [Unsloth runs 552B DeepSeek-V4.1-Flash-1017 locally at 1-bit in about 125 GB using SSD offloading](https://unsloth.ai/docs/new/changelog#deepseek-v4.1-flash-1017-and-flux.3)
2. [AWS adds push-based KV transfers to vLLM for up to 3x better time-to-first-token](https://aws.amazon.com/blogs/machine-learning/push-based-kv-transfer-to-improve-ttft-and-tpot-for-disaggregated-inference-in-vllm/)
3. [Prime Intellect's prime-rl trains Qwen3-235B-A22B 1.15x faster than torchtitan on identical hardware](https://www.primeintellect.ai/blog/prime-rl-0-10-verifiers-0-4)
4. [UltraBERT-Embed-150M embeds up to 32K tokens into one 768-dimensional vector](https://qdrant.tech/blog/ultrabert-embed/)
5. [OpenCode pairs a full-access coding agent with a read-only planning agent](https://github.com/anomalyco/opencode) · 70 Stars today
6. [HandsOn benchmarks robotic manipulation with 66 tasks and 18,000 human demonstrations](https://huggingface.co/papers/2610.26765)