Analysis 3 min read

Three Answers

Three Answers

Kimi K3, DeepSeek V4 Pro, and Qwen3-Coder-Next all shipped open weights this summer. One weighs 1.6 terabytes. One just raised API prices 355%. One runs on a MacBook.

Same label. Three different products.

What Each Optimized For

The open-weight coding model space didn't converge this summer. It diverged.

Capability Frontier

Kimi K3

2.8T params · 1.56 TB on disk
93.4% SWE-bench Verified
(Vals AI independent)
#1 Frontend Code Arena
To run it
8+ H100 GPUs · ~$200K hardware
“Open” means
Auditable. Not accessible.
Algorithmic Efficiency

DeepSeek V4 Pro

1.6T params · 49B active (MoE)
80.6% SWE-bench Verified
(vendor reported)
Codeforces 3206
To use it
Their API
$3.96/M output (was $0.87)
“Open” means
Weights available. Access managed.
Deployment Efficiency

Qwen3-Coder-Next

80B params · 3B active (MoE)
70.6% SWE-bench Verified
(paper + community verified)
Runs on consumer GPUs
To run it
RTX 5090 or 64GB MacBook
Hardware you might own
“Open” means
Download. Run. Own the inference.

Moonshot wanted to close the gap to the proprietary frontier. It worked — K3 sits 1.6 percentage points below Claude Fable 5 on SWE-bench Verified, independently confirmed by Vals AI. But the weights are 1.56 terabytes. Running them requires data center hardware — eight H100 GPUs minimum. The model that proved open weights can reach the frontier also proved that the frontier has nothing to do with accessibility.

DeepSeek wanted the best reasoning per parameter. V4 Pro achieves Codeforces 3206 and 80.6% SWE-bench Verified. For three months, it paired that with the cheapest frontier-class API in the market. That ended four days ago.

Alibaba asked a different question entirely: what runs on hardware a developer already owns? Qwen3-Coder-Next — 80 billion total, 3 billion active — achieves 70.6% on SWE-bench Verified on a single consumer GPU. Apache 2.0 license. No API. No pricing page.

The Pricing Floor

For three months, DeepSeek V4 Pro charged $0.87 per million output tokens. At that price, the algorithmic efficiency niche doubled as the accessibility niche — best reasoning per parameter AND cheapest to use. It looked like the model that would commoditize inference.

On August 16, DeepSeek switched to peak/off-peak billing:

Before Aug 16
$0.87/M output
Flat rate · all hours
After Aug 16
$3.96/M peak
$1.98/M off-peak

A 355% increase at peak. Off-peak is still more than double the old rate. DeepSeek described it as “allocating resources more reasonably.”

The repricing revealed what loss-leader pricing obscured: algorithmic efficiency doesn’t mean cheap. It means efficient use of expensive resources. “Best per token” and “cheapest per million” were always different claims. The first is architectural. The second was a strategy.

Developer reaction was predictable: audit token usage, test alternatives, set up multi-provider routing. Some teams are now looking at Qwen3-Coder-Next and a $2,000 GPU.

I’ve tracked the commodity thesis for ten articles (#32→#70). Models are commodities, value is in scaffolding — that claim broke at the restricted frontier but held in the accessible tier. What the last month shows is that the accessible tier didn’t commoditize. It specialized.

Three models, three optimization targets, three relationships to the word “open.” The capability frontier is open by license — you can audit the weights, but you need $200,000 in hardware to run them. The efficiency niche is open by weight availability — the MIT-licensed model exists on disk, but most users interact through an API priced at the vendor’s discretion. The deployment niche is open by practice — you download 40 gigabytes, plug in a GPU, and own the inference.

Only one of these is what most people mean when they say “open.”