Strata 0.2.12 Adds Big Speed Gains and Agent Tool Calls

Developer Niko Veit's latest release targets budget GPUs, AMD and Apple hardware, plus a batch of agent tool-call fixes.

The latest release of an open-source project called Strata claims you can run a frontier-class model on ordinary consumer hardware. Solo developer Niko Veit dropped version 0.2.12 today, and the pitch is blunt: a graphics card with just 12 GB of VRAM is enough.

The model in question is Qwen3.8-Flash-Next from Alibaba's Qwen team. It's a mixture-of-experts design with about 180 billion parameters in total. Only 125 billion of those actually do work on any single token. The rest sit in system RAM and on your SSD, computed in parallel while your GPU handles the cached portion. You don't need a pricey card to run it.

Keep in mind that this is where the real trick lives. Strata keeps the most-used experts in VRAM and computes the long tail on the CPU at the same time. Neither side waits on the other. As the project's README states, "a cheap RTX card can run 128B models. Your GPU VRAM no longer caps what AI model you can run. You just need system RAM."

Bold claim. This is a project barely two months old.

As of early November 2026 it reports about 41,300 GitHub stars and 4,500 forks, so plenty of people are watching. The first version shipped on September 24, and it has moved at what can only be called a frightening pace. One independent tracker counted "two weeks, 44 releases" and the model itself barely changed. Most of the work is engine speed and hardware support. For what it's worth, that velocity is exactly what some users love and others are quietly terrified of, which says plenty about open-source AI in late 2026.

The headline wins, by hardware

The release leans hard on benchmark data, and not without reason. Veit's team ran 0.2.11 next to 0.2.12 on identical machines, several rounds each, swapping which version ran first to dodge bias. The headline number comes from a 12 GB GPU. On an RTX 3060-class card, prompt reading jumped from 87 to 264 tokens per second at 2K context, and from 361 to 981 at 32K. If you're keeping score, that's roughly a threefold speedup.

Bigger or differently shaped chips saw smaller gains, which is entirely expected. AMD's Strix Halo desktop chips ran 1.24x to 2.14x faster on two quantizations. A Mac with an M5 Max picked up 10.3 to 12.7 percent on output and 11.7 to 13.9 percent on prompts. The Radeon R9700 added 8.6 to 11.6 percent on long prompts.

A tuned drafter, then a reversal

The biggest structural change is a re-tuned helper model called DFlash2, which now turns on automatically on more hardware. It's a small download, about 0.34 GB, though Intel Arc users get a 0.52 GB version. Strata keeps its own tuned copy.

Auto now pulls DFlash2 on Mac with two quantizations, NVIDIA cards with all experts resident, the DGX Spark, Strix Halo, Intel Arc and 12 GB cards.

But here's the interesting part. On AMD's gfx1201 hardware, which covers the R9700 and RX 9070, 11 of 50 DFlash2 engines froze during prompting on the previous build. So Strata flips those cards back to its own MTP drafter by default. The on-screen message tells you exactly what happened, and you can still force DFlash2 with strata launch --drafting dflash2 if you want to gamble. It's a "trust but verify" stance, and frankly a refreshing one.

The agent angle: tool calls fixed

Strata isn't only a chat interface anymore. It's positioning itself as an offline backend for coding agents like Claude Code, Codex CLI and OpenCode, so this release spends real energy on tool calling.

The most cited fix is #4828, a community pull request by a contributor named gdevenyi. It took forced tool selection from working in just 0 of 24 runs to working in all 24. Another change, #4832, makes Anthropic tool results follow the correct reference ID. There are XML-tag parsing fixes too, plus a new setting that controls how calls to undeclared tools are handled.

How to get it

The release ships 42 files totaling 9.70 GiB. You can pull it with the standard install script, or the PowerShell variant on Windows. Then launch it, or point a coding tool at it:

curl -fsSL https://github.com/Niko1221/Strata/releases/latest/download/install.sh -o install.sh && sh install.sh
strata launch
strata connect opencode --write

Model files aren't part of the Strata repo, and they arrive from third parties with their own licenses. Budget for storage too, since the roughly 28.8 GB n-gram table lives on your SSD.

Known problems are listed openly. Two-card peer mode can fault on its first request in about 4 percent of fresh starts. AMD gfx1201 users may still see DFlash2 stall if they force it. And there are measured losses: JSON extraction on an RTX 3060 dropped 18.9 percent. Strata is candid at least.

Head here to the full release notes if you want every number and every SHA-256 hash.