Software 45260 Published by

Strata 0.2.19 just launched with native support for Thinking Machines Lab's Inkling 2 Mini, a 26.4-billion-parameter "looped transformer" that reasons longer internally by re-running its middle layers. The model fits on hardware as small as a 4 GB graphics card through three labeled quantization packs ranging from 10.47 GB to 19.35 GB, and Strata claims it reaches up to 250 tokens per second on an RTX 5090. The release also adds NVIDIA Blackwell FP4 support and broad Qwen speed boosts while continuing the project's from-scratch kernel rewrite, all with zero environment variables required. It arrives hours after 0.2.18 made Tencent's 220-billion-parameter Hy4-Flash run on a single PC, reinforcing Strata's reputation for fast iteration and unflinchingly honest quality labeling.



Strata 0.2.19 ships a "reasoning" model that runs on anything, the day after it made a 220-billion-parameter model fit on a desk

One day after shipping Tencent's Hy4-Flash, Slovenian developer Niko Veit pushes a 26.4-billion-parameter "looped transformer" onto hardware as small as a 4 GB graphics card.

Hours after Strata made a 220-billion-parameter model run on a single PC, it did something smaller and stranger. Version 0.2.19, released today, adds native support for a new reasoning model that Thinking Machines Lab calls Inkling 2 Mini. It is only 26.4 billion parameters. It also runs on a graphics card with just 4 GB of memory.

Screenshot_from_2026_10_24_14_26_18

On whatever you've got

Inkling 2 Mini is a "looped transformer." Most models apply each layer once per word. This one runs its two middle blocks three times each, so 24 stored layers become 48 applications per token. The idea is not new. Universal transformer research suggested the middle layers carry the most abstract reasoning, so re-running them lets a model compute longer internally before it says anything aloud.

Thinking Machines measured that three passes roughly doubled a model's no-chain-of-thought time horizon on one benchmark. A variable-loop research variant kept improving even at ten loops, twice as many as it was trained with.

On the lab's own benchmarks Inkling 2 Mini leads every open-weights model under 100 billion parameters. It hits 48.2% on DeepSWE v1.1 and 46.5% on the Vals Index v2.2, ahead of comparably sized models like Gemma 4 31B and Qwen3.8-27B. It even launched with a video of itself building a playable flight simulator offline on a MacBook Pro in airplane mode.

Keep in mind those are the lab's numbers, and they come from the people who built the model. Still, a flight simulator that runs offline on a laptop is a memorable demo.

The quantization packs are labeled with how closely they match the original. The smallest, SQ-10, costs 10.47 GB on disk and targets 12 GiB cards or 16 GiB Macs, running from just 4 GiB in hybrid RAM plus VRAM mode at 92.35% agreement. SQ-15 costs 14.35 GB and lands at 96.68%. SQ-20 is the full build at 19.35 GB, bit-exact and matching 100% of the time.

That is a far cry from Hy4-Flash, whose packs ranged from 130 to 181 GB each. Inkling fits the size of a movie because the model itself is an order of magnitude smaller.

Strata's notes are unusually precise about speed. On an RTX 5090, Inkling 2 Mini reaches up to 250 tokens per second in the SQ-10 build. That figure climbs past 4,089 tokens per second across 64 concurrent requests, showing how speculative decoding turns one fast card into a small serving cluster. An Intel Arc Pro B70 hits 159 tokens per second, an Apple M5 Max lands near 114, and an old GTX 1080 Ti still produces about 24.6.

It's a rather impressive spread for a model from one developer, though I would hold the headline figures loosely. Those peak speeds come from developer-tested high-end rigs. The realistic floor for most machines sits closer to 64 GiB of RAM than the advertised 12 GiB of VRAM.

An engine, not a wrapper

Strata 0.2.19 rounds out a quiet consolidation. It started replacing llama.cpp kernels back at 0.2.14, and by 0.2.16 its own NVIDIA, AMD, Intel and Metal kernels ran with no ggml underneath. Now four model families run through purpose-built shape engines: the original Qwen and DeepSeek, the Hy4-Flash from 0.2.18, and now Inkling.

The throughline is that each new model forced new kernels. Looping needed a merge between passes. Hybrid attention needed its own sparse paths. The idea of keeping busy experts in VRAM and streaming the rest from RAM became reusable everywhere.

Strata says its own kernels produce token-for-token identical output to reference implementations. SQ-20 matches Inkling bit-exactly with zero divergence, and the speculative-decoding path adds none either.

This answers the "is this really different from llama.cpp?" question raised against 0.2.18, structurally rather than politely. It is shipping four bespoke engines, one of them a from-scratch rewrite.

The release cycle is the real story here. 0.2.18 landed on 21 November. 0.2.19 on 22 November. The 0.2.x series runs on roughly a two- to three-day cycle. Two days earlier, 0.2.17 flipped a speed shortcut off after users reported worse answers. A public retraction of a default. Then, two days later, a brand-new architecture shipped with full hardware coverage.

What is most durable, though, is the honesty. Strata still publishes a known-limits section and ships each pack labeled with its exact agreement percentage. In an era of black-box benchmarks, showing your work is a small discipline worth keeping.

Where to get it

The desktop installers are available at GitHub, roughly 300 MB for Linux, 109 MB for Windows, and 161 MB for Mac, all signed where applicable. Terminal users can run the install script and then strata launch, or run strata update for future upgrades. The app downloads 10 to 19 GB per pack on first use, so cap your output tokens and keep an eye on memory.

Head here for the full release notes.