The download counter on Hugging Face read 4,000. One week after Thinking Machines dropped Inkling-Small โ the model that supposedly rivals China's best open-weight architectures โ the market responded with a shrug. That number is the first hard datum in a story buried under a mountain of benchmark hype. But the real signal isn't in the downloads. It's in the pricing page. And the pricing page tells a different story than the press release.
Let me be precise about what we're looking at. Thinking Machines, founded by former OpenAI CTO Mira Murati, released Inkling-Small as the compact variant of a larger, unreleased Inkling model. The architecture is textbook Mixture-of-Experts: 276 billion total parameters, 12 billion activated per token. The benchmarks are impressive on their face โ SWE-Bench Verified at 80.2%, Terminal Bench 2.1 at 64.7%, AIME at 95.1% under max-effort sampling. The context window stretches to one million tokens, with native multimodal support and a 256K-token serverless API. The positioning is clear: a US-built, open-weights model that can go toe-to-toe with DeepSeek and Qwen while offering the compliance pedigree that Western enterprises supposedly crave.
But here's the unspoken catch: the pricing math doesn't hold up under scrutiny. The company claims Inkling-Small costs "about half" of OpenAI Luna. The actual numbers: $0.30 per million input tokens and $1.20 per million output tokens. OpenAI Luna sits at $0.20 input and $1.20 output. That's not half. That's 50% more expensive on input and identical on output. The "half" narrative only works if you assume a heavily input-dominant workload โ a convenient assumption that doesn't match real-world agentic use patterns, where output tokens dominate. This isn't a rounding error. It's a framing choice that tells me exactly which benchmarks and which pricing comparisons Thinking Machines wants you to see.
The deeper story emerges when you stack Inkling-Small against its actual competitive set. DeepSeek V4-Flash โ the model that quotes for everything โ runs at $0.14 input and $0.28 output. That's four times cheaper on output. Kimi K3 from Moonshot charges $3.00 and $15.00 respectively, which makes Inkling-Small look modest, but Kimi's pricing reflects a different market position. The honest comparison is against DeepSeek, because both are open-weights models targeting the same developer base. And on that axis, Inkling-Small loses on cost by a wide margin. Its only structural advantages are the multimodal capabilities DeepSeek lacks, the million-token context window, and the made-in-USA supply chain. Those are real advantages. The question is whether they justify a 4.3x output premium. The pool remembers what the ticker forgets: open-weight markets are brutally price-sensitive at the developer tier.
Now let's talk about what the benchmarks actually measure. SWE-Bench Verified at 80.2% is a strong score โ top-tier in the real world as of early 2025, where the best models hover in the 70-75% range. But here's the problem: the article doesn't specify the sampling strategy. In my experience auditing model claims, "max effort" or "best-of-n" settings can inflate scores by 5-10 percentage points. AIME 95.1% with max effort is impressive, but it's not the same as a single-pass evaluation. Terminal Bench 64.7% suggests real capability in terminal-driven agentic tasks, which has profound security implications that we'll get to shortly. The absence of evaluation protocol details doesn't mean the scores are fabricated โ it means I can't verify them, and neither can you. Volatility is the tax on uncertainty, and the uncertainty here is substantial.
What's missing from this entire announcement is more revealing than what's present. No training FLOPs. No GPU hours. No grid size. No training methodology. No alignment details. No red-teaming disclosures. No model card transparency. For a company selling to enterprises that "care about provenance, supply chain, and regulatory consistency," the silence on training data provenance and safety evaluation is deafening. In 2017, I audited ICO whitepapers for a living. I learned to spot the difference between projects that omitted details because they were commercially sensitive and projects that omitted details because the details wouldn't survive scrutiny. This has the texture of the latter.
The strategic logic, however, is coherent. Thinking Machines is running a three-layer commercialization play: open weights on Hugging Face for developer acquisition and trust-building; a serverless Tinker API for low-friction revenue and usage data; and a fine-tuning API at $1.73 per million tokens with a 50% intro discount to lock developers into custom-weight ecosystems. This is a classic land-and-expand strategy. The fine-tuning hook is the smartest part โ once a developer builds business-specific weights on top of Inkling-Small, switching costs become prohibitive. Code is law, but audits are mercy; the same logic applies to developer ecosystems. Once you're locked in, you're locked in.
But the economics of that fine-tuning price point deserve scrutiny. $1.73 per million tokens for fine-tuning is priced as if it were inference. Fine-tuning costs are dominated by training time, not token throughput. This is either a loss leader designed to buy market share โ which tells me the market share isn't coming naturally โ or a deliberately misleading metric that obfuscates actual GPU-hour costs. Either way, it's a sign of cold-start pressure. Companies don't offer 50% discounts when they have more demand than supply. Speculation is just data with a heartbeat, and the data here says Thinking Machines is hungry.
The geopolitical layer is the most interesting part. For years, the frontier open-weight race has been dominated by Chinese labs โ DeepSeek, Moonshot, Alibaba's Qwen. American labs kept their best models behind closed APIs. Thinking Machines is the first serious US attempt to compete on open weights directly. Mira Murati's founder story makes this credible; she was instrumental in shipping ChatGPT, she knows how to build product, and she has the network to raise capital. But here's the contrarian angle that everyone's missing: the "US supply chain" differentiator is not just a feature, it's a confession. It acknowledges that Chinese open models are technically competitive but commercially unreachable for Western enterprises with regulatory constraints. Inkling-Small is designed for that gap. The problem? The US government doesn't have a supply chain problem โ it has a cost problem. American compute and labor are structurally more expensive. No amount of engineering optimization can close a 4x cost disadvantage if the underlying factors โ electricity, GPUs, salaries โ are all higher. Entropy increases until someone audits it, and the entropy here is structural.
Now let me raise the security question that should be keeping every CISO awake. These benchmarks that Thinking Machines is so proud of โ particularly Terminal Bench at 64.7% โ mean this model is genuinely capable at terminal command execution. That's a dual-use capability. Open weights mean anyone can download and use it, bypassing whatever safety filters the API layer might have. There's no mention of jailbreak resistance, no red-team results, no safety evaluation of dangerous capabilities. For a model with agentic software engineering and terminal operation skills, this is not a minor oversight. It's the kind of oversight that leads to front-page headlines. Based on my audit experience, I'd say this: the security documentation is not just incomplete, it's inconsistent with the enterprise sales pitch. You can't sell to a bank on compliance provenance while withholding the safety data that bank's security team needs to make a procurement decision. Liquidity doesn't lie, and neither does institutional pushback โ wait until the first enterprise security review gets a look at this model card.
The competitive matrix is equally revealing. Against DeepSeek, Inkling-Small loses on price but wins on multimodal and context length. Against Kimi, it wins on price but loses on ecosystem maturity. Against Meta's Llama, it's competing on the same turf โ open weights, US supply chain โ but with a fraction of the ecosystem support. The most damning comparison is against Llama, because the "first US frontier-level open-weight model" narrative falls apart the moment you remember that Meta's Llama family exists. It's not the first. It's not even the most downloaded. The 4,000 first-week downloads are roughly two orders of magnitude below what a successful open-weight launch looks like in 2025. The truth is hidden in the gas fees โ or in this case, in the Hugging Face download counter.
Let me also flag the benchmark naming anomaly. "AIME 2026" โ if this is meant to be the annual American Invitational Mathematics Examination, the 2026 version wouldn't exist in the article's presumed 2025 timeline. This could be a codename, a forward-looking reference, or a factual error. When I saw it, I didn't just raise an eyebrow. I re-read the entire technical section looking for other inconsistencies. The vague comparison to a "four-times-larger model" with no named baseline is another red flag. Four times larger than what? Inkling? DeepSeek? Claude? The unwillingness to name the baseline is a rhetorical choice that serves no one except the marketer who needs to avoid direct comparison with the strongest closed-source models.
The investment picture is a function of narrative, not numbers. Murati's pedigree commands premium valuation in today's AI market โ I've seen comparable founder stories raise at multi-billion valuations on little more than a slide deck. Inkling-Small provides product-market evidence for the story. But the 4,000 downloads, the lack of disclosed revenue, the absence of enterprise customer names, and the missing API usage data all point to a company still in cold-start. The valuation will be driven by what investors imagine, not by what the data shows. Rewriting the rules before the bug writes them โ that's what good founders do. But the bug here might be the pricing model, and it needs serious auditing before investors write the next check.
Here's what I'm watching now. The fine-tuning ecosystem is the only credible moat โ if developers start building meaningful custom weights on Inkling-Small, the switching costs compound. But that requires a developer base that the download numbers don't yet support. Next, watch for the Inkling model, the full 975-billion-parameter version. If Thinking Machines can ship a close-sourced frontier model with top-tier scores, the calculus changes. And watch the enterprise deals โ the company claims to target compliance-conscious organizations, so a single public deployment announcement from a bank or defense contractor would be worth more than 10,000 Hugging Face downloads. Until then, the pool remembers what the ticker forgets: open-weight competition rewards the patient, and the patient are waiting.
The bottom line: Inkling-Small is a solid MoE model with competent architecture and competitive capabilities in specific areas. But it is not the open-weight revolution it claims to be, and it is not going to displace DeepSeek on cost or performance-per-dollar. Its only real moat โ the US supply chain and regulatory alignment โ is a political asset, not a technical one. And in a market that moves on measurable quality and price, politics only gets you the first meeting, not the contract.
You think open-weight competition is about benchmarks and token prices? Look at the download counters. Look at the fine-tuning discounts. Look at what gets disclosed and what gets buried. The code is public, but the context is not. And the context is where the real battle for the next phase of AI infrastructure is being fought.