local AI

Oct
02
Banner showing Pi 1.0 connecting a local model server, such as Ollama or llama.cpp, with files, shell commands, and MCP tools through Codemode. Pi Durable is marked experimental.

What is the Pi coding agent, and how does it work with local models?

Pi 1.0 shipped on October 1 with built-in MCP support through Codemode, after Earendil publicly rejected MCP. Here's how to run it against Ollama or llama.cpp, which packages you need, and when the experimental Pi Durable framework fits.
8 min read
Sep
26
Invide GLM-5.3 banner with a chart comparing GLM-5.3, Kimi K3, DeepSeek V4 Pro and Flash by score and cost per index task; below, GLM-5.3 leads Flash 42% to 33% on Terminal-Bench 4.0.

GLM-5.3 benchmarks, pricing and how it compares with Flash and rivals

GLM-5.3 keeps GLM-5.2's list price and matches Kimi K3's cost per Artificial Analysis index task. DeepSeek V4 Pro costs a third as much on the same index. Existing users must update requests that disable thinking.
8 min read
Sep
17
GLM-5.3-Flash model guide comparing its $0.50-per-million output price with Opus 4.8 at $25, alongside questions about cost per finished task and its 306 GiB FP8 weights.

GLM-5.3-Flash costs up to 50x less than Opus. Your cost per finished task may not.

Given a billion tokens per vulnerability, GLM-5.3-Flash matched Claude Mythos Preview on ExploitBench at roughly 6% of the cost, while one production team found it fine for bounded work and not ready for their hardest tier.
10 min read
Sep
15
DeepSeek V4.1 Flash banner showing a 552B MoE backbone, 8B active on input, 16B on output, a 1M-token context, and $0.003/M off-peak cached input.

DeepSeek V4.1 Flash is a 552B model that bills like a smaller one

DeepSeek V4.1 Flash activates 8B parameters per input token and prices cached input from $0.003 per million. On Terminal-Bench 4.0, it scores 31.2 to Opus 5.0's 51.8.
6 min read
Aug
26
Apple M5 Ultra Mac Studio banner comparing 512GB memory with 1.2TB/s bandwidth and noting that 512GB requires the 36-core/80-core chip.

Apple's M5 Ultra Mac Studio bets on bandwidth for local AI

Apple kept M5 Ultra's 512GB ceiling but raised bandwidth to 1.2TB/s. The 256GB model starts at $9,499; prompt-prefill gains remain untested.
3 min read
Aug
12
News banner for Meta Muse Glimmer showing Zuckerberg’s “The Future is for Everyone” essay alongside the released Glimmer weights and the still-promised Muse Spark 1.2 weights.

Zuckerberg says AI should be for everyone. Muse Glimmer is Meta’s first test

Muse Glimmer is a real 30B Apache 2.0 release, but its performance on common hardware, independent benchmark results, and Meta’s promised follow-ups remain open.
3 min read
Aug
04
MiniMax H3 banner showing Base weights released, while Context-IR and 2K regeneration remain unavailable locally under a restricted license.

MiniMax H3’s weights are public, but the full video system is not

MiniMax released H3 Base weights with local ComfyUI workflows, but its restricted license, large downloads, and missing 2K modules limit practical use.
3 min read
Aug
03
Qwen3.8-Max banner showing its hosted API available now, while open weights for Max and the 27B model remain promised for next week.

Qwen released its new Max model, but the open weights are still a promise

Qwen released Qwen3.8-Max and promised open weights for it and a new 27B model next week, but developers still need the artifacts, license, and local tests.
2 min read
Jul
15
PrismML Bonsai 27B benchmark summary showing 3.9 GB binary and 7.2 GB ternary builds, with 11 tokens per second on an iPhone.

Bonsai 27B fits on a phone. But does it run well?

PrismML compressed a 27B-class model into a phone-sized package, but kernels, memory overhead, and real device performance still determine whether it is practical.
3 min read