AI / Models · what shipped, checked · 2 tracked

Models

A hand-checked model registry — licence, weights availability with the date we checked, prices with the date they held, and benchmarks that say who measured them.

A row of machined steel blanks of the same shape, one in copper, on paper.

DeepSeek-V4-Pro-0813

DeepSeek · released 2026-08-13

Parameters
1.7T (mixture-of-experts)
Context
1M tokens
Max output
384K tokens
Licence
MIT
Weights
Yes — Hugging Face (checked 2026-08-20)
Price in / M
$0.66
Price out / M
$1.98
Cached / M
$0.022
Prices as of
2026-08-20
Providers
12

Increase took effect 2026-08-16.

No independent benchmark in the registry yet — vendor tables are deliberately not reproduced here.

What changed, and does it matter

The open-weight release the harness question was waiting for: an MIT-licensed 1.7T MoE with a swappable-plugin harness alongside it. The pricing moved up three days after launch — the meter is already being tuned. No independent benchmark has landed in the registry yet; vendor tables are deliberately not reproduced here.

Qwen3.8-27B

Alibaba (Qwen) · released 2026-08-14

Parameters
27.78B (marketed 27B)
Context
262,144 native (~1M via YaRN)
Max output
not separately capped
Licence
Apache 2.0
Weights
Yes — Hugging Face (ungated) (checked 2026-08-21)
Price in / M
$0.43
Price out / M
$3.1
Cached / M
—
Prices as of
2026-08-21

Hosted price; the consumer-hardware story is the point of this model.

Benchmarks — who measured what

  • Artificial Analysis Intelligence Index52independent · Artificial Analysis
  • 24GB-GPU throughput (5.01 BPW quant, Q4 KV)50.4 tok/s mean over ten runsindependent · piszczek.pl (personal blog, 2026-08-17)
  • RTX 5090 task suite (Q4_K_M)14/16 · ~94 tok/sindependent · KGP Talkie (2026-08-14)

What changed, and does it matter

A hybrid decoder — 64 layers but only 16 full attention, the rest linear-attention Gated DeltaNet — which is why its KV cache is a quarter of what the layer count suggests. That arithmetic is the difference between "cannot possibly fit in 24GB" and "fits with room to spare", and it is the subject of the sizing tutorial this entry backs.

Model notes / what the table can't say

The prose

No model notes published yet. Back to the hub.