Public story · 2026-03-19 · source-backed

Mamba-3: Inference-First SSM Beats Transformers 4%, Runs 7x Faster.

Issue

2026-03-19

Confidence

source-backed

Sources

Redaction

redacted

Story

Together.ai released Mamba-3 (Apache 2.0, ICLR 2026) — an SSM achieving ~4% better language modeling than the Transformer baseline while running up to 7x faster on long sequences. Key innovations: Exponential-Trapezoidal Discretization, Complex-Valued SSMs with the RoPE Trick, and MIMO decoding for higher hardware arithmetic intensity. The architecture is explicitly inference-first, targeting agentic workloads where inference — not training — is the bottleneck. Together.ai

Source trail

Together.ai

Entities

Apache Complex Exponential Faster First SSM Beats Transformers ICLR Inference Mamba MiMo RoPE Trick Runs SSM Together Transformer Trapezoidal Discretization Valued SSMs

Provenance

Canonical issue: Ramsay Research Agent — 2026-03-19
AI generated: no
Story unit: 2026-03-19-mamba-3-inference-first-ssm-beats-transformers-4-runs-7x-faster
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

Mamba-3: Inference-First SSM Beats Transformers 4%, Runs 7x Faster.

Story

Related stories

Mamba-3 Beats Transformers on Decode Latency at 1.5B Scale.

Gemma 4: 1,200% Improvement in Agentic Tool Use.

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

Reduced Matrix Multiplication cuts Transformer inference with no weight changes, and attention is far more reducible than MLPs.

Cohere open-sourced a 30B coding model that runs on a single H100. Agentic coding without the frontier bill.

Two Labs Ship Sub-500MB Agentic Models in the Same Week. Edge AI Agents Just Became Real.

Mistral Small 4: 119B MoE Unifies Instruct, Reasoning, Vision, and Coding

MUD Extends Muon Optimizer to Full Transformer Architecture.

Source trail

Entities

Provenance

Mamba-3: Inference-First SSM Beats Transformers 4%, Runs 7x Faster.

Story

Related stories

Mamba-3 Beats Transformers on Decode Latency at 1.5B Scale.

Gemma 4: 1,200% Improvement in Agentic Tool Use.

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

Reduced Matrix Multiplication cuts Transformer inference with no weight changes, and attention is far more reducible than MLPs.

Cohere open-sourced a 30B coding model that runs on a single H100. Agentic coding without the frontier bill.

Two Labs Ship Sub-500MB Agentic Models in the Same Week. Edge AI Agents Just Became Real.

Mistral Small 4: 119B MoE Unifies Instruct, Reasoning, Vision, and Coding

MUD Extends Muon Optimizer to Full Transformer Architecture.

Source trail

Entities

Provenance