The Signal — July 20, 2026
Three stories today that all circle the same question: when a lab or company makes a big claim, how much of it can you actually check? A Chinese giant unveils a trillion-parameter model with no benchmarks attached, a rocket company reportedly angles its way into defense compute, and a small lab ships something you can download and verify yourself.
Alibaba shows off Qwen3.8-Max, and asks you to take the numbers on faith
At Shanghai's World AI Conference on July 19, Alibaba's Qwen team unveiled Qwen3.8-Max-Preview, a 2.4-trillion-parameter multimodal model it describes as "second only to Fable 5" among currently available systems. A preview build went live the same day on Alibaba's Token Plan, the Qoder IDE, and the QoderWork desktop assistant at roughly 10% of standard pricing, with a promise that full open weights are coming "soon."
What Alibaba has not shipped is anything you could use to test that framing: no model card, no license, no active-parameter count, no per-token price for the full model, and no independent benchmark. The comparison to Anthropic's Fable 5 is Alibaba's own, and early hands-on reports have been mixed, with users calling the preview token-heavy. The timing is the real story. This lands days after Moonshot's Kimi K3, and the two biggest Chinese labs are now in an open sprint over who can put a credible frontier model in developers' hands first. Parameter counts make headlines. Downloadable weights and reproducible scores are what will actually settle it.
Sources: MarkTechPost · Startup Fortune
SpaceX is reportedly in talks to sell the Pentagon AI compute
The Wall Street Journal reported on July 17 that SpaceX and the Defense Department are in early-stage talks over a multibillion-dollar arrangement to supply dedicated AI data-center capacity. According to the reporting, that capacity would be earmarked for divisions including the National Security Agency and for personnel deploying AI in the field. Reuters and Bloomberg picked up the story, though neither SpaceX nor the Pentagon confirmed it, and the WSJ's sourcing is people familiar with the matter rather than anyone on the record.
For now this is an unconfirmed report about talks, not a signed contract. If it holds up, though, it points somewhere worth watching. A company better known for rockets would be moving directly into defense AI infrastructure, and it would be doing so as a compute supplier undercutting the likes of CoreWeave. Anthropic already tapped SpaceX's Colossus facility in Memphis for 300 megawatts earlier this year, so the compute ambitions are real even if this particular deal stays a rumor for now.
Sources: Wall Street Journal · Reuters via CNA
Feyn's SQRL checks the database before it writes the query
Feyn Labs offers a counterweight to the day's unverifiable claims, and it's one you can download and test. The lab released SQRL, a family of text-to-SQL models that run read-only probes against a database before committing to an answer. The idea addresses a real failure mode. A query can be valid SQL and still be wrong, joining the wrong tables or filtering for values that do not exist, and none of that throws an error. By inspecting the actual data first, SQRL can resolve ambiguities that a schema alone would hide.
The flagship SQRL-35B-A3B reports 70.6% execution accuracy on the BIRD Dev benchmark, edging Claude Opus 4.6's 68.77% by Feyn's own measurement, and it distills into 4B and 9B checkpoints that keep most of the quality at a fraction of the size. That benchmark number is self-reported, so the usual caution applies, but the weights are public on Hugging Face, which means anyone can check the claim rather than take it on faith. In a week of trillion-parameter announcements, a self-hostable model built around the boring discipline of looking before it leaps is a quietly useful thing.
Sources: MarkTechPost · Feyn Labs · Hugging Face
On the Editor's Desk
A three-story Signal today. We looked at a handful of others and set them aside. NVIDIA's DeepStream 9.1 and Moonshot's Kimi K3 were both strong, but we already covered them, and there's no new angle to add yet. A peer-reviewed study linking AI use to higher resilience among finance workers was genuinely interesting, but it rests on a single journal with no independent corroboration, so it needs more before we put it in front of you. And a couple of policy items, including the FTC's proposed rule on AI output steering, are still stuck in comment periods with nothing new to report.
