Infra2026-06-12 · source-backed

Zyphra's ZAYA1-8B was trained from scratch on AMD Instinct, not NVIDIA.

Issue

2026-06-12

Confidence

source-backed

Sources

Redaction

redacted

Story

Per a roundup, ZAYA1-8B is an Apache-2.0 sparse-MoE with 8B total params and ~760M active per token, trained entirely on AMD hardware. The AMD-only training run is the signal: the open-weight training stack is diversifying off NVIDIA. Verify the numbers against the official model card before relying on them, but the hardware story alone is worth noting.

NVIDIA invested in OpenAI / Shared entities / Shared topic / What happened next
Meta ships Muse Glimmer, a 30B open agent model
Linked by a graph relationship (NVIDIA invested in OpenAI); both cover AMD, Apache, MoE, NVIDIA; overlapping topics (against, model).
NIM built by NVIDIA / Shared entities / Shared topic / What happened next
Local models became a first-class feature of hosted coding agents in a 48-hour window
Linked by a graph relationship (NIM built by NVIDIA); both cover Apache, MoE, NVIDIA; overlapping topics (model, nvidia).
NVIDIA invested in OpenAI / Shared entities / Shared topic / What happened next / Tension
AMD may put $5B into Anthropic against a 2-gigawatt MI450 commitment.
Linked by a graph relationship (NVIDIA invested in OpenAI); both cover AMD, Nvidia; overlapping topics (against, instinct, nvidia).
Cursor partners with NVIDIA / Shared entities / Shared topic / Earlier coverage
NVIDIA Nemotron 3 Nano: Open Agentic Model with RL Training Pipeline
Linked by a graph relationship (Cursor partners with NVIDIA); both cover MoE, NVIDIA; overlapping topics (active, model, nvidia, training).
NVIDIA released Blackwell / Shared entities / Shared topic / Earlier coverage / Downstream implication
llama.cpp Gets Backend-Agnostic Tensor Parallelism, and the NVIDIA Tax Just Got Optional
Linked by a graph relationship (NVIDIA released Blackwell); both cover AMD, NVIDIA; overlapping topics (card, model, nvidia).
Alibaba partners with NVIDIA / Shared entities / Earlier coverage
DeepSeek V4 Preview Drops: 1.6T Parameters, 1M Context, Apache 2.0, and a 35x Price Gap That Should Make Every Builder Rethink Their Inference Budget
Linked by a graph relationship (Alibaba partners with NVIDIA); both cover Apache, MoE, NVIDIA; earlier Apache coverage from 2026-04-24.
Hugging Face partners with NVIDIA / Shared entities / What happened next
llama.cpp now runs a 295B model on one machine
Linked by a graph relationship (Hugging Face partners with NVIDIA); both cover Apache, MoE, NVIDIA; picks up the Apache thread on 2026-07-15.
NVIDIA benchmarked against DeepSeek-R1 / Shared entities / Shared topic / Earlier coverage
Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE
Linked by a graph relationship (NVIDIA benchmarked against DeepSeek-R1); both cover Apache, MoE; overlapping topics (active, model).

Source trail

roundup

Entities

AMD AMD Instinct Apache MoE NVIDIA Verify ZAYA1 Zyphra

Provenance

Canonical issue: Ramsay Research Agent — June 12, 2026
AI generated: no
Story unit: 2026-06-12-zyphra-s-zaya1-8b-was-trained-from-scratch-on-amd-instinct-not-nvidia
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

Zyphra's ZAYA1-8B was trained from scratch on AMD Instinct, not NVIDIA.

Story

Related stories

Meta ships Muse Glimmer, a 30B open agent model

Local models became a first-class feature of hosted coding agents in a 48-hour window

AMD may put $5B into Anthropic against a 2-gigawatt MI450 commitment.

NVIDIA Nemotron 3 Nano: Open Agentic Model with RL Training Pipeline

llama.cpp Gets Backend-Agnostic Tensor Parallelism, and the NVIDIA Tax Just Got Optional

DeepSeek V4 Preview Drops: 1.6T Parameters, 1M Context, Apache 2.0, and a 35x Price Gap That Should Make Every Builder Rethink Their Inference Budget

llama.cpp now runs a 295B model on one machine

Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE

Source trail

Entities

Provenance

Zyphra's ZAYA1-8B was trained from scratch on AMD Instinct, not NVIDIA.

Story

Related stories

Meta ships Muse Glimmer, a 30B open agent model

Local models became a first-class feature of hosted coding agents in a 48-hour window

AMD may put $5B into Anthropic against a 2-gigawatt MI450 commitment.

NVIDIA Nemotron 3 Nano: Open Agentic Model with RL Training Pipeline

llama.cpp Gets Backend-Agnostic Tensor Parallelism, and the NVIDIA Tax Just Got Optional

DeepSeek V4 Preview Drops: 1.6T Parameters, 1M Context, Apache 2.0, and a 35x Price Gap That Should Make Every Builder Rethink Their Inference Budget

llama.cpp now runs a 295B model on one machine

Gemma 4 Drops Four Open Models, and the 31B Dense Beats a 397B MoE

Source trail

Entities

Provenance