Strata's v0.1.43.1 hotfix pulls a speed regression and patches crashes for its local 125B model engine
Strata, the open-source engine that lets you run a 125-billion-parameter AI model on an ordinary gaming PC, shipped v0.1.43.1 today. It's a hotfix for the version released the day before. And it exists because that earlier release went out the door with a couple of defaults that made things slower, badly slower on some machines, on real hardware.
The headline fix takes adaptive overlap back to opt-in everywhere. In v0.1.43 the maintainers switched that optimization on automatically for NVIDIA-on-Linux setups. It helped a few cards. On others it wrecked decode speed.
An RTX 5070 Ti desktop lost 21 percent. The RTX 5060 Ti lost 19. A RTX 5090 Laptop under WSL2 lost 21. Only one RTX 5090 desktop gained a marginal 1 or 2 percent.
Since Strata can't test every possible hardware combination, the change is now opt-in globally. You set STRATA_ADAPT_OVERLAP=1 to turn it on, and the startup log now tells you whether it's running and why it couldn't be. Keep in mind that the answers don't change: outputs stay byte-identical whether you flip the switch or not.
On Linux NVIDIA desktops, decode speed had collapsed to single digits in the worst cases. An RTX 4090 with a Ryzen 9800X3D was decoding at 19 to 37 tokens per second when it should have been pushing 102 to 116. The culprit was a thread quirk: the engine pins its decode thread to one core, and a CUDA callback thread inherited that same core, adding roughly a millisecond of wait time per layer.
The fix warms up that thread first and drops the dependency in the common path. On an RTX 3060 test, decode went from 36.5 to 56.5 tok/s. That's a couple of seconds shaved off a long reply. Worth having.
There were hangs too. Long prompts on Windows could stall at "layer 1 of the prompt chunk" and restart the whole engine. On SATA SSDs they ran roughly seven times slower than in v0.1.41. The startup check had switched to reading the n-gram table through mmap, which tripped a 60-second watchdog churning through rows one page fault at a time from a cold 27 GB table. The hotfix stops the auto-switch and keeps direct reads. On a 77,800-token prompt that's 884 tok/s versus 76 to 154 when mapped.
Note the 15,000-rows-per-second performance bar was set on NVMe drives. A real SATA SSD tops out around 10 to 14K rows/s with direct reads, so the watchdog was basically expect hardware to do something it can't.
Two cards, older AMD cards, and Intel all get repaired
Two NVIDIA cards running with --batch were about 30 percent slower than in v0.1.41. Automatic batch groups had collapsed into one serial group spanning both cards. The hotfix restores pipelined groups, and a 2× RTX A4000 test went from 62.3 to 85.2 tok/s combined. That's roughly the 35-percent throughput restoration Strata calls out in the notes.
RX 6000 cards (gfx103x) on Windows had seen prompts drop 16 to 20 percent since v0.1.40.2. An RX 6800 test climbed from 584 to 596 up to 798 to 808 tok/s once the draft layer's side stream was recreated.
Intel two-card splits had a nasty edge case: miss an expert in VRAM and the first answer could vanish, leaving the engine pointed at a device that no longer existed. Setup now writes STRATA_STAGE_TRIM=1 for two Intel cards so every expert stays resident.
An RX 7900 XTX doing multi-slot serving with images crashed after about 30 minutes under load, an issue traced to host-to-GPU copies still routing through pageable memory. The hotfix routes those through a ring of pinned buffers, verified against a 65-minute multi-client stress test.
There's exactly one genuine behavior change here, and it's oddball. A user message that quoted </tool_call> was previously triggering something. Now it's read as plain text.
On the server side, a batch slot that returned HTTP 400 "BGEN: no such free slot" could leave a conversation broken. The server now waits for or takes another slot and returns a retryable 503 with a Retry-After header instead. Two servers sharing one GPU could both dump their model onto a single card if a before_load hook failed; that path now checks free VRAM and either loads with a warning or bounces the request.
Why a 125B model on a gaming PC still matters
Running a 125B-parameter model locally used to mean a server rack and a real electricity bill. Strata's approach treats your GPU, RAM, and SSD as one shared memory pool. The README likens it to a kitchen: the ingredients you grab most stay on the counter, everything else waits in the pantry.
The model in question, Qwen3.8-Flash-Next, is a mixture-of-experts design. Of its 125 billion parameters, only about 6 billion activate per token. A small multi-token-prediction layer drafts up to three next tokens and the full model verifies them, netting a 1.6 to 1.8× speedup while producing byte-identical output. The engine claims 50 to 120 tok/s on consumer hardware, faster than most people read short replies.
That latency is arguably competitive with a paid API for short answers. Whether it competes on accuracy is a separate question, and the self-reported benchmarks pitting it against Claude-Opus-class models deserve a healthy grain of salt. Independent replication is thin, and the project labels some figures as estimates.
The hardware reality bites hard. You need 12 GB or more of VRAM and 32 to 64 GB or more of system RAM. And it's the RAM that actually decides which model variant fits, Coder at 32 GB, IQ2_XS at 48 to 64 GB, IQ3_S at 64 GB or higher. A ~4-bit Unsloth variant needs roughly 94 GB and reads from the SSD unless your RAM is ample.
Some casual observers have been blunt about it. The cost of "125B on consumer hardware" isn't the GPU, they say, it's 64 GB of DDR5, which can outcost the graphics card. Strata is free under the MIT License. The one-click installer detects your graphics card and downloads the roughly 70 GB model.
Head here to pull v0.1.43.1 from the project's GitHub repository.
