Strata 0.2.12 is out for users running language models locally. It speeds up reading long inputs on some GPUs and Macs and fixes errors in requests that invoke external tools. Users with 12GB GPUs can spend less time waiting for input processing. Developers integrating tools can more reliably handle requests that specify one of several tools.
1. Tuning the auxiliary model that predicts upcoming words
The key change is an in-house tuned DFlash2 drafter, an auxiliary model that helps generate output by predicting the main model's next words.
Automatic mode now selects it on more devices, including Macs, Strix Halo, Intel Arc A-series GPUs and some Nvidia cards. On Macs, it is the default auxiliary model for both supported quantized models, stored in reduced size.
On 12GB Nvidia cards, the tuned version is automatically applied to UD-IQ4_XS-v31. On Linux, a fix stops counting memory already used by the engine as free space, reducing fallbacks to the slower file-cache path. Input acceleration includes this memory-handling change as well as auxiliary model selection.
2. Input throughput, the amount processed over time, roughly tripled on a 12GB card
The developers compared 0.2.11 with the new candidate 2–3 times on each device. They restarted the engine and alternated the order, generating 300-token responses by always choosing the highest-probability token. Each version used its saved device-specific settings, reflecting the new auxiliary model selection.
The developers report that UD-IQ4_XS-v31's 2K input throughput on an RTX 3060 12GB rose from 87 to 264 tokens per second. For 32K inputs, the same pairing rose from 361 to 981 tokens per second. The changes reduce input-reading wait times for both short and long inputs.
On Strix Halo running Windows, the developers report that UD-Q4_K_XL's 4K input throughput rose from 597 to 1,267 tokens per second. This setup also enabled two separate optimizations: fused prompt kernels for UD-Q4_K_XL and the Q5_0 layers of UD-IQ4_XS-v31, and grouped decode kernels.
On the Mac M5 Max, the developers report that UD-Q4_K_XL's 4K input throughput rose from 1,406 to 1,605 tokens per second. Output speed for the same model increased by 12.7% across tested workloads using a geometric mean, which reflects relative changes. The Mac input-processing path now overlaps CPU work with GPU operations.
3. Fixing requests that specify one of several tools
The developers report that tests forcing a specific tool choice with multiple tools declared improved from 0 successes in 24 attempts to 24 in 24. The fix directly addresses requests that failed to execute even when told to use a tool.
Model names sent by Claude Desktop now map to the currently loaded model.
More tool-call opening tags are now recognized. For MCP, the external tool connection protocol, a tool name given only by its final portion is matched if exactly one tool has that suffix.
Anthropic-format tool results are linked to their original requests using call identifiers, the values that distinguish tool calls.
Undeclared tool calls are returned as text by default, but users can opt to receive them as calls.
4. Output gains and losses varied by workload
The developers report the following output changes by workload for the RTX 3060 12GB with UD-IQ4_XS-v31. Users focused on code generation and agents, programs that select tools to perform tasks, saw gains in output as well as input processing.
Table: Output speed changes by workload on the same card and model
Conditions: RTX 3060 12GB, UD-IQ4_XS-v31, 0.2.12 candidate versus 0.2.11, 2–3 runs, engine restarted each run, alternating version order, 300 tokens generated by highest-probability token selection, per-version device settings
Unit: %
| Workload | Output speed change |
|---|---|
| Code generation | +37 |
| Agents | +34 |
| JSON extraction | -18.9 |
The developers report that UD-IQ4_XS-v31 tool calling on Strix Halo slowed by 9.5%. They attributed this measured loss to the tuned auxiliary model.
These comparisons are results for the new candidate, not measurements repeated on the final engine package. The subsequent engine source changes were limited to the version number and two build fixes for older Nvidia cards.
5. Update in the app or terminal
Update in the app or run the following command in a terminal. The app and terminal versions are 0.2.12, while the control and engine packages are 0.1.58.
strata updateThe tuned version downloads once, when automatic mode first needs it. Automatic selection reduces manual setup but adds a 0.34GB download for the standard file or 0.52GB for Intel Arc. Without internet access, it uses the standard auxiliary model or the existing prediction method. Manually selecting the existing prediction method prevents downloading the tuned version.
Strata 0.2.12 improves input wait times and tool connections; adoption should depend on performance changes in frequent workloads rather than peak speed.