Libre Computer

Changing the Landscape of Open Computing with reliable, composible, upstream, and secure SBCs and Computing Devices for professionals, hobbyists, and tinkerers.

[Kimi.ai @Kimi_Moonshot]

Introducing Kimi K3: Open Frontier Intelligence

🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost
🔹 Built for long-horizon agentic coding and self-evolving workflows

Kimi K3 is now live on on https://t.co/zrk6zZxZUo, Kimi Work, Kimi Code, and the Kimi API.
Open Weights by July 27, 2026.

🔗 API: https://t.co/XCrgjXAqMw
🔗 Tech blog: https://t.co/YTfiMSNM1f

[Libre Computer @librecomputer]

AOSP is coming.

[Libre Computer @librecomputer]

@sbccompare Every codec supported including encoding, faster graphics, 2D acceleration, video pipeline improvements, and more.

[Libre Computer @librecomputer]

@geerlingguy Dense ~30B models (needing 40GB VRAM) is where local inferencing becomes useful. GB10's 256b DDR inferface is the bottleneck that prevents it from being useful. RTX Pro 6000 is the real starting point. This may change with model advancements a year from now.

[Libre Computer @librecomputer]

@will_whang We have an ASUS with Snapdragon X Elite. What a nightmare machine.

[Libre Computer @librecomputer]

@cnxsoft We expect to release everything May 1st. This is just the tip of the iceberg.

[Libre Computer @librecomputer]

@FanlessTech @raspberrytorpes We are still here.

[Libre Computer @librecomputer]

@geerlingguy Can you refer? We made significant progress on the software side so we want to reach out to them.

[Libre Computer @librecomputer]

@aniiland @ASUS @Snapdragon We use it for native compiles and now this part of our infrastructure is down. We moved to a Dimensity 9400 machine in meantime.

@Qualcomm would capture so much of the local inference market with proper Linux support of Hexagon with the X2.

[Libre Computer @librecomputer]

@ASUS we have one of your @Snapdragon X Elite S5507QAD for our testing environment. The battery discharged to 0% and became stuck at 0%. As hardware engineers, we opened the battery and tried to send SMBUS commands to reset the lock. No luck on such a trivial problem. Lenovo and Dell can both reset from Windows/BIOS. What are you doing?

[Libre Computer @librecomputer]

@harukaze5719 The world need 10x the LP5 lines to keep up at the rate designs are scaling.

[Libre Computer @librecomputer]

@ValdikSS We haven't gotten to Rockchip and Allwinner optimizations yet. Those will probably land in Q3.

[Libre Computer @librecomputer]

@luigifcruz La Frite is $30 retail. More powerful than a Pi 3.

[Libre Computer @librecomputer]

@ValdikSS Tritium?

[Libre Computer @librecomputer]

@ValdikSS It was a longer run than the others, 5 mins vs 1 mins. There's automatic temperature throttle hooked in.

[Libre Computer @librecomputer]

Together with draw call optimization, our first platform can achieve up to 5x draw call improvement and 2x shader throughput improvement from Debian 13's mesa baseline.

[Libre Computer @librecomputer]

@fahdmirza They need to fit it into their 16GB B50.

[Libre Computer @librecomputer]

@lone_voidwalker @cnxsoft LP core performance are not insignificant. You can fit 3-4 LP cores in the same space as HP cores. There are a lot of tasks where discrete LP cores running two threads provide much higher performance than one HP core sharing resources.

[Libre Computer @librecomputer]

@Yuchenj_UW Add a few 0s.