A local creative engine and a local reasoning model may look like two separate launches. To us, they are the same idea: the best work should not have to leave your Mac.
Today we are bringing LTX 2.5 and Qwen 3.8 27B into Kynd. One expands what a local studio can make. The other expands how confidently a local agent can think, use tools and keep a long piece of work moving.
“LTX 2.5. Qwen 3.8 27B. Both shipping in Kynd. Keep the same character across a story, then put a serious local reasoning model behind the team producing it. No remote queue. No API key. Just your Mac, doing more.”
Your local film finally gets a recurring cast.
Local video has been brilliant at moments. The harder problem is continuity: the same person, prop or place surviving the cut into the next shot. Our new LTX 2.5 Ingredients path is built for that problem.
Give Kynd a small set of local reference images or video frames and it composes them into a reusable Ingredients sheet. That reference conditions the native LTX render, so a character is more than a sentence we hope the model interprets the same way twice. It becomes an asset the production can deliberately carry from shot to shot.
Reusable production ingredients
Characters, locations, props and visual motifs live on the board as named assets, ready to attach to the shots that need them.
Direct reference conditioning
The reference sheet reaches the patched native MLX Serve path with the official Ingredients adapter—without inventing an unnecessary starting frame first.
Mac-aware rendering
Kynd checks the runtime, memory tier, adapter and input files before unloading chat for a render. If the complete video and soundtrack do not arrive, the result is not quietly accepted.
Three reusable inputs. One coherent shot.
LTX 2.5 · frame sequence
ResultThe character, guitar and warm camper interior arrive in the generated motion—not just in the prompt.
Consistency turns a striking clip into a story.Kynd Film · LTX 2.5 Ingredients
Ingredients is available as an explicit, experimental quality tier at launch. The UI tells you when the patched runtime and official adapter are ready, and it tells you when they are not. That distinction matters on a local system: “supported” should mean the whole path is present on this Mac, not that a checkbox exists somewhere.
Qwen 3.8 27B, tuned for the loop.
Qwen 3.8 27B arrives as a dense hybrid model with vision, a large native context window and model-authored reasoning controls. Kynd maps its familiar Quick, Standard and Deep choices directly onto the model’s native low, medium and xhigh effort levels.
More importantly, Kynd preserves the model’s reasoning state through multi-turn tool work. The model can inspect a file, call a tool, absorb the result and continue from the same line of thought. It does not have to cosplay a fresh assistant after every tool call.
The fastest context is the context you do not prefill twice.
A capable model is only half the experience. Agent work repeats an enormous prefix: system instructions, tools, conversation and the work completed so far. Sending all of that back through the model on every turn makes a local agent feel slower as it becomes more useful.
Kynd’s fork caching saves the model’s processed state at a safe conversation boundary. On the next turn, it loads that fork and prefills only the new material. Each active conversation gets its own slot, so switching tasks can restore the right cache instead of making two conversations evict each other.
We also warm stable system prefixes when the server is genuinely idle. And while a tool is running, Kynd can extend a separate shadow fork with the tool output. The original cache remains untouched until the next prompt proves the extension correct at the token level.
Our published conversation-fork measurements were made on an earlier Qwen 3.6 build, not Qwen 3.8: a 20K-character prefill fell from 11.2 seconds to 0.3 seconds. That number is evidence for the cache architecture, not a promise that every model and every Mac will reproduce the same timing.
One product, several honest local routes.
For Qwen 3.8 27B, the launch profiles cover native MLX 8-bit and high-fidelity Q8 GGUF through llama.cpp. More broadly, Kynd’s model layer can route across the local backends below. The model picker shows which route a pack uses before you load it.
The principle is simple: a backend badge is not a guarantee of equal speed. MLX is the preferred Qwen 3.8 route on Apple Silicon. GGUF is the compatibility route. Kynd calibrates context, sampling, caching and concurrency to the actual combination instead of flattening every local model into “localhost.”
Shipping means integrated.
LTX 2.5 is wired into the Film board, its Ingredients shelf and the render lifecycle. Qwen 3.8 27B is wired into model detection, memory tiers, native reasoning effort, vision and the agent’s cache rules. Both come with explicit readiness checks and safe fallbacks.
That is the release: not two model names in a dropdown, but two new local capabilities made useful by the rest of Kynd.