A $387 mini PC, the Morefine M700S, sits on a shelf next to Intel and AMD boxes and runs a processor whose instruction set shares not one line of code with x86 or Arm. That is the Loongson 3A6000, and in May 2026 the company confirmed it had shipped a million of them. A million units is a rounding error next to what Intel moves in a week. For a chip built on an architecture the rest of the planet does not use, running a software ecosystem that exists almost entirely inside one country, a million is a real number, and it is the cleanest evidence yet that the most stubborn bet in Chinese silicon is starting to pay off.
LoongArch is that bet. Every other company in this series rents its instruction set from somewhere: Hygon and Zhaoxin pay for x86; HiSilicon and Phytium build on Arm. Loongson designed its own ISA from the ground up and licenses nothing from the West. The lineage runs back to the Godson project at the Institute of Computing Technology under the Chinese Academy of Sciences, which started in 2001 under chief architect Hu Weiwu, was originally derived from MIPS, spun out as a commercial entity in 2010, and floated on Shanghai’s STAR Market in 2021. Hu summed up the whole strategy in one sentence at the 2025 product launch: China cannot build a house on someone else’s wall forever. That is the company in fourteen words, and it explains both why Loongson is the most strategically important name on this list and why it loses money doing it.
The desktop chip that quietly reached Zen 3
The 3A6000 is the part that earned the headlines, and it deserves them. It runs Loongson’s LA664 core, with four cores and eight threads, simultaneous multithreading, clocked between 2.0 and 2.5 GHz, and posts SPEC CPU2006 base scores around 46.1 integer and 57.7 floating-point. Translated out of benchmark-speak, that puts it near an Intel Core i3-10100F and roughly at Zen 3 levels of instructions per clock. Compared to the previous 3A5000, it gained more than 60% on single-core integer and around 90% on floating-point. Reaching Zen 3 IPC matters far more than the raw position in the stack, because Loongson did it on an instruction set it invented, with a compiler ecosystem it had to build itself, with none of the decades of tuning that Intel and AMD inherited. The honest caveat is the part nobody in the press release wants to dwell on: it is still a four-core chip in a world where a midrange Ryzen ships sixteen, so a six-core Ryzen 5 will run rings around it in anything multithreaded. For the government desktop, the school lab, the tax office terminal it was actually built for, four fast-enough cores are plenty.
The server chip nobody outside China saw coming
On June 26, 2025, in Beijing, Loongson launched the 3C6000, and this is the chip that moves the company out of the schools-and-tax-offices ghetto and into the data center. It uses the same LA664 microarchitecture as the 3A6000 but scales it through a chiplet design wired together by Loongson’s own LoongLink interconnect, in three flavors: a single-chiplet 16-core part, a dual-chiplet 32-core part, and a quad-chiplet 64-core part. Built on a 12nm process, the 3C6000 posts SPEC CPU2017 scores around 7.3 single-thread and 55 multi-thread, and in its 64-core configuration Loongson benchmarks it in the neighborhood of Intel’s Xeon 8380 on integer work, with the 16-core version compared to a Xeon Silver 4314. It carries Level-2 certification from the China Information Technology Security Evaluation Center, the top tier for mission-critical deployment, and feeds four channels of DDR4-3200. China Telecom’s eCloud signed on as a showcase deployment. A 64-core domestic server CPU benchmarking against a Xeon 8380, on a wholly Chinese ISA, would have sounded like fantasy three years ago.
| Chip | Cores / threads | Node | Performance reference | Target |
|---|---|---|---|---|
| 3A6000 | 4C / 8T | mature node | ~Core i3-10100F, Zen 3 IPC | Desktop, government |
| 3C6000 (S/D/Q) | 16 / 32 / 64C | 12nm | 64C ~Xeon 8380 integer | Servers, cloud |
| 2K3000 / 3B6000M | industrial / mobile | mature node | +256 GFLOPS GPU, 8 TOPS AI | Edge, embedded |
| 3A7000 (roadmap) | desktop | 7nm | +20-30% over 3A6000 | Desktop |
| 3D7000 (roadmap, ~2027) | >32C chiplet | sub-10nm | DDR5, PCIe 5.0 | Servers |
Below the headline parts sits the 2K3000 and 3B6000M, aimed at industrial control and mobile computing on the smaller LA364E core. What makes them interesting is the integration: Loongson folded in its own second-generation GPU at around 256 GFLOPS plus an AI block rated at 8 TOPS, which is the company quietly extending from pure CPUs into the heterogeneous compute everybody else is chasing. It is not going to trouble a Snapdragon, but it means Loongson now has a graphics and AI story that is entirely in-house, on an ISA it controls, which is the recurring theme with this company. Everything is theirs, for better and for worse.
The roadmap, and the node ceiling
What comes next tells you where the real constraint lives. The 3A7000 moves to a 7nm process for a claimed 20 to 30% uplift and, per reporting, will be produced entirely inside China because of sanctions. The 3D7000, penciled for around 2027, is a sub-10nm chiplet design with more than 32 cores, DDR5, and PCIe 5.0, with Loongson saying the IP design work has already started. The 3D5000 stitches two 3C5000 dies into a 32-core part that scales to 128 cores in a four-way server. Notice what every one of these has in common: the architecture keeps advancing on schedule while the manufacturing node lags years behind TSMC’s leading edge, because Loongson, like everyone else in China, is downstream of what SMIC can actually deliver. The design team is not the bottleneck. The fab is. That is the single sentence that describes the entire Chinese CPU industry, and Loongson is just the purest case of it because it has nothing else to blame.
Money that does not add up, on purpose
The financials are where it gets genuinely strange. Loongson’s 2024 revenue came in around 504 million yuan, roughly $70 million, essentially flat year over year, while the net loss widened to about 625 million yuan from 329 million the year before. Trailing twelve-month revenue sits near $75.9 million against a market capitalization around $8.85 billion. Run that math and you get a company valued at well over a hundred times revenue while losing money and shrinking its industrial-control segment by 44%. No sane investor prices a chip company that way on its fundamentals. The market is not valuing Loongson as a semiconductor business at all. It is pricing a strategic option on Chinese instruction-set sovereignty, a bet that if the day ever comes when Arm and x86 are fully weaponized against Beijing, the one company that owns its ISA outright becomes priceless overnight. The bright spot underneath the loss is real, though: information-technology chip revenue jumped over 190%, carried by the 3A6000 and a recovering e-government procurement cycle.
Sanctioned for the one thing it didn’t do
Here is the irony I cannot get past. In March 2023 the US Commerce Department added Loongson to the Entity List, citing support for the People’s Liberation Army’s modernization. Loongson is the only company in this entire series that licenses no Western instruction set, the one firm that by design cannot be choked off through Arm or x86, and it got sanctioned anyway. The export controls aimed at the others have an obvious lever, the ISA, and Loongson deliberately removed that lever years before anyone thought to pull it. So Washington reached for the only thing left, the manufacturing equipment and the American tools inside SMIC, which is the same wall every Chinese designer hits. The sanction is almost a backhanded compliment: there was nothing clever to cut, so they cut the wafers. Loongson, meanwhile, has a quiet side market in Russia and Belarus, where Western sanctions already pushed buyers toward anything non-American, and a captive pipeline at home that dwarfs both. A single school-district trial in Hebi put 10,000 Loongson PCs into classrooms, and the stated ambition is a program that could put tens of millions of Loongson machines into Chinese schools every year by 2030.
That captive demand is also the thing I keep circling back to as the unanswered question. Loongson survives today because the Chinese state guarantees it buyers through the Xinchuang domestic-substitution program, through schools, ministries, and state telecoms that are told to buy domestic and increasingly told to buy a fully independent ISA. The software desert is the price of admission: choosing LoongArch means porting every application, maintaining a Linux world that exists nowhere else, and living without Windows entirely. The open question is whether any of that captive volume ever matures into demand. Loongson wins on merit, in an open market, against a Ryzen that costs the same and runs everything. I do not think we will know for years, and I am not convinced Loongson needs the answer to be yes. A company that owns the one instruction set Beijing can never have revoked is valuable to the state, whether or not it ever turns a profit, and that, more than any benchmark, is why the most uncommercial chipmaker in China is also the one with the most secure future.
This is one of five company deep dives in a series on China’s fabless CPU makers. Start with the overview.