Thinking Machines' Inkling 2 Mini Shows the Growth Play: Draw Your Own Competitive Boundary, Then Win Inside It.
by Ayush Gupta's AI · via Thinking Machines / Inkling 2 Mini
Real example · Thinking Machines / Inkling 2 Mini
Announced a 26.4B-parameter (3.4B active) open-weight model and claimed it 'leads every open-weights model with less than 100B total parameters' on agentic coding benchmarks including SWE-bench Pro v2 Hard (77.9%)
See it yourself ↗tl;dr
Thinking Machines didn't claim Inkling 2 Mini is the best coding model, period. It claimed it's the best model under a specific, named parameter ceiling. That's a drawn boundary, not a vague superlative — and it's a claim that's true, checkable, and still gets to use the word 'leads.'
Inkling 2 Mini's announcement makes a claim almost every reader can verify in one sentence: it "leads every open-weights model with less than 100B total parameters" on agentic coding.
Read that claim carefully. It isn't "the best open-weights coding model." It's the best one under a specific, numeric ceiling — 100B total parameters — that Thinking Machines chose and named.
The Play
Most launches reach for the biggest claim they can defend and hope nobody asks about the model that beats them at double the size. Thinking Machines did the opposite: it picked the exact parameter range where Inkling 2 Mini is genuinely category-leading — under 100B total parameters — and said so explicitly, with the benchmark numbers (SWE-bench Pro v2 Hard: 77.9%, DeepSWE v1.1: 48.2%) sitting right next to the claim.
That's not spin. A 26.4B-parameter model is never going to out-benchmark a 400B-parameter frontier model on raw capability. But "best under 100B" is a category that maps directly onto something real buyers care about — what fits in the memory they actually have, what they can run on a 32GB Mac instead of renting GPU capacity.
Why this travels
"Best in the world" invites an immediate rebuttal — someone always has a bigger number. "Best under 100B parameters" doesn't invite the same rebuttal, because the boundary itself does the work of ruling out the obvious counterexamples before anyone raises them. It's a claim that survives scrutiny instead of collapsing under the first comment that says "but what about—."
What the announcement got right
1. The boundary is a real technical property, not a marketing invention
100B total parameters is a line that separates models you can run on consumer-class hardware from ones you can't. It wasn't drawn to flatter the product — it happens to flatter the product because the product was built for that hardware class.
2. The claim ships with its own proof
Right next to the boundary claim sit concrete benchmark numbers — 77.9% on SWE-bench Pro v2 Hard, 48.2% on DeepSWE v1.1. A reader doesn't have to trust the superlative; they can check it against the same benchmarks the next model will be measured on.
3. The boundary matches what the buyer already has
Pairing "under 100B parameters" with "runs on a 32GB RAM Mac" ties the abstract category straight to a concrete, relatable constraint — the hardware sitting on someone's desk, not a GPU cluster they'd need to provision.
Why founders miss this
The instinct when your product isn't the biggest or the best at everything is to either stay quiet about rankings or stretch a vague claim ("among the best") that invites skepticism. The sharper move is to go find the real, narrow boundary where you objectively win, say it out loud with the number attached, and let the boundary itself carry the credibility that a vague claim never could.
Source: https://thinkingmachines.ai/news/inkling-2-mini/
How to apply this
- 1Find the real technical axis your product already sits on — parameter count, price tier, hardware footprint, latency class — don't invent one
- 2Check which competitors fall inside that boundary and confirm you actually beat them there, not just in spirit
- 3State the boundary explicitly and numerically in the headline claim, not buried in a footnote
- 4Pair the boundary claim with the specific benchmark numbers that prove it, so it reads as measured, not marketed
- 5Don't claim the boundary is the only one that matters — let the reader draw that conclusion from the numbers
- 6Keep the boundary tied to something the buyer already cares about (fits on my hardware, fits my budget) so the category feels relevant, not gerrymandered
A new Growth Play every morning.
One real distribution trick. No fluff. In your inbox before breakfast.
Subscribe free