DEV Community

AI Field Notes Series' Articles

Back to ai maya's Series
The great sparsification: 2026 open-weight MoE models now activate 1 to 9 percent of their weights
Cover image for The great sparsification: 2026 open-weight MoE models now activate 1 to 9 percent of their weights

The great sparsification: 2026 open-weight MoE models now activate 1 to 9 percent of their weights

Comments
6 min read
When the attacker is an agent: a defender's field guide to autonomous AI intrusions in 2026

When the attacker is an agent: a defender's field guide to autonomous AI intrusions in 2026

Comments 1
6 min read
Decision AI Is Becoming Its Own Category, and the Metrics Are Not Ready

Decision AI Is Becoming Its Own Category, and the Metrics Are Not Ready

Comments
3 min read
Ling-3.1-flash and the 7:1 hybrid: how much KV cache does a 560B model need at 1M tokens?
Cover image for Ling-3.1-flash and the 7:1 hybrid: how much KV cache does a 560B model need at 1M tokens?

Ling-3.1-flash and the 7:1 hybrid: how much KV cache does a 560B model need at 1M tokens?

Comments
10 min read
Can two Mac Studios really run an 816B model? A bandwidth roofline check on Muse Spark Open
Cover image for Can two Mac Studios really run an 816B model? A bandwidth roofline check on Muse Spark Open

Can two Mac Studios really run an 816B model? A bandwidth roofline check on Muse Spark Open

Comments
12 min read
Shorter outputs, same price: what DeepSeek V4.1-Flash-1017 changes in an agent's bill
Cover image for Shorter outputs, same price: what DeepSeek V4.1-Flash-1017 changes in an agent's bill

Shorter outputs, same price: what DeepSeek V4.1-Flash-1017 changes in an agent's bill

Comments
13 min read
Is open-weight really 90% as good at 10% of the cost? Checking IBM's claim against list prices
Cover image for Is open-weight really 90% as good at 10% of the cost? Checking IBM's claim against list prices

Is open-weight really 90% as good at 10% of the cost? Checking IBM's claim against list prices

Comments 1
12 min read
Inference NPUs That Train: What a 325M Pretraining Run on a Rebellions ATOM-Max Shows, and What It Does Not
Cover image for Inference NPUs That Train: What a 325M Pretraining Run on a Rebellions ATOM-Max Shows, and What It Does Not

Inference NPUs That Train: What a 325M Pretraining Run on a Rebellions ATOM-Max Shows, and What It Does Not

Comments
2 min read
Mistral Large 4 on your own GPUs: does 1 trillion parameters fit on 8 H200s, and what does the license allow?
Cover image for Mistral Large 4 on your own GPUs: does 1 trillion parameters fit on 8 H200s, and what does the license allow?

Mistral Large 4 on your own GPUs: does 1 trillion parameters fit on 8 H200s, and what does the license allow?

Comments
13 min read
Kimi K2 closes in December: what moving to K3.1 Pro changes, and what the default reasoning effort costs
Cover image for Kimi K2 closes in December: what moving to K3.1 Pro changes, and what the default reasoning effort costs

Kimi K2 closes in December: what moving to K3.1 Pro changes, and what the default reasoning effort costs

Comments
12 min read
Marvell is buying Lumentum for $115B: what the optical bet means for AI cluster power, and what a fixed 4.5x exchange ratio really locks in
Cover image for Marvell is buying Lumentum for $115B: what the optical bet means for AI cluster power, and what a fixed 4.5x exchange ratio really locks in

Marvell is buying Lumentum for $115B: what the optical bet means for AI cluster power, and what a fixed 4.5x exchange ratio really locks in

Comments
12 min read
DeepSeek's 160,000 Huawei Ascend chips: what 10,000 chips for weeks actually buys, and why the reports disagree on training
Cover image for DeepSeek's 160,000 Huawei Ascend chips: what 10,000 chips for weeks actually buys, and why the reports disagree on training

DeepSeek's 160,000 Huawei Ascend chips: what 10,000 chips for weeks actually buys, and why the reports disagree on training

Comments
11 min read
Anthropic's FlashProve: what 72x overhead and a 1,024-token cap mean for proving which model answered you
Cover image for Anthropic's FlashProve: what 72x overhead and a 1,024-token cap mean for proving which model answered you

Anthropic's FlashProve: what 72x overhead and a 1,024-token cap mean for proving which model answered you

1
Comments 1
10 min read
Baidu's ERNIE-5.5-Flash-Base: can you self-host a 280B open model, and who is allowed to?
Cover image for Baidu's ERNIE-5.5-Flash-Base: can you self-host a 280B open model, and who is allowed to?

Baidu's ERNIE-5.5-Flash-Base: can you self-host a 280B open model, and who is allowed to?

Comments
12 min read
Inkling 2 Small is cheap per token but writes a lot: what 150,000 output tokens per task does to the price
Cover image for Inkling 2 Small is cheap per token but writes a lot: what 150,000 output tokens per task does to the price

Inkling 2 Small is cheap per token but writes a lot: what 150,000 output tokens per task does to the price

Comments
12 min read
Anthropic's $3.4B gas generator deal: what $2,000 per kW buys, and what the reports leave out
Cover image for Anthropic's $3.4B gas generator deal: what $2,000 per kW buys, and what the reports leave out

Anthropic's $3.4B gas generator deal: what $2,000 per kW buys, and what the reports leave out

Comments
11 min read
Anthropic's IPO filing, line by line: how the $900.9B compute figure breaks down, and what $84.6B covers
Cover image for Anthropic's IPO filing, line by line: how the $900.9B compute figure breaks down, and what $84.6B covers

Anthropic's IPO filing, line by line: how the $900.9B compute figure breaks down, and what $84.6B covers

Comments
12 min read