C.W.K.
Stream
Lesson 05 of 05 · published

Production Deployments — and Honest Limitations

~14 min · production, deployments, limitations

Level 0Observer
0 XP0/50 lessons0/14 achievements
0/100 XP to next level100 XP to go0% complete

What's actually in production

Mamba-family architectures are now in real production at multiple major labs:

AI21 Jamba 1.5 — 256K context, available on AWS Bedrock, Azure, GCP. The first Mamba-family model to hit major cloud marketplaces. Its 1.5 release was the moment hybrid-Mamba moved from research to enterprise procurement.

NVIDIA Nemotron-H 8B / 56B / 47B — 92% attention-replaced-by-Mamba-2. About 3× faster than Llama-3.1 70B at 65K context. The 56B was pretrained in FP8 on 6,144 H100s over 20T tokens — production scale by any measure.

IBM Granite 4.0 — 9:1 Mamba:attention ratio, 70%+ RAM reduction relative to comparable pure Transformers. IBM's enterprise customers see a smaller serving footprint at comparable quality.

TII Falcon Mamba 7B — beats Llama-3.1 8B on standard benchmarks while being a pure SSM (interesting precisely because it's pure — proves selectivity gets you most of the way at 7B scale).

Cartesia Llamba-8B — 12× throughput vs. its Llama 3.1 8B teacher, achieved by distilling a Transformer into a Mamba student. The distillation route is increasingly important: instead of training a Mamba from scratch, take a strong existing Transformer and convert it.

The honest limitations don't go away

Production validation doesn't repeal physics. Mamba-family models still have:

  • Five-shot MMLU gap — pure SSMs noticeably underperform on few-shot in-context learning. Hybrids close most of this; pure SSMs do not.
  • Asymmetry bias from nonlinear convolution — Mamba has a slight bias toward tokens earlier in the sequence due to how the selective scan accumulates information.
  • Associative recall failures — structural, as established in the previous track.
  • Narrow learning rate windows — production teams have to spend more on hyperparameter sweeps relative to Transformer recipes.
  • Tooling gap — improving fast, but still measurably behind Transformer in 2026.

None of these are deal-breakers for the right workload. All of them are reasons to default to a Transformer (or a Transformer-heavy hybrid) when the workload doesn't specifically reward Mamba's strengths.

External links

Exercise

Pick one Mamba-family production model (Jamba 1.5 Mini is a friendly choice — open weights, well-documented). Pull it from Hugging Face, run it locally, generate against your own long-context test (a 50K-token document of your choosing) and compare quality to Llama 3.1 8B on the same input. Note specifically where the Mamba wins (memory, throughput) and where it loses (questions requiring exact retrieval of named entities deep in the document).

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.