Models we train.
Recipes we run.
Karti-Small models for support, vision, and agent tool calling. The full catalog includes the evidence, release status, and practical recipes for serving Qwen on DGX Spark.
Karti-Small models
Four public model pages. The results include the regressions.
Karti-Small-Support-9B
Policy-grounded support replies and actions, trained with a reward computed in code.
Released LoRA adapter
0.531 → 0.706 held-out composite
32 episodes. Verification asks regress from 5/7 to 2/7.
Karti-Small-VL-4B
Read screens and panels. Decline when an identifier is unreadable.
Released BF16 v1 weights
0.718 → 0.907 on ScreenSpot-v2
1,271 held-out examples. General vision stays approximately flat.
Karti-Small-VL-4B-NVFP4
A smaller quantization of VL-4B v1, with the vision tower preserved.
Released Blackwell build
21.1 → 51.8 output tok/s
Single stream on DGX Spark, 512 output tokens; some accuracy is traded for speed.
Karti-Small-Agent-3B
A SmolLM3-based agent program that observes state, calls tools, and delivers through a tool.
Card and recipe published · weights pending
Train/serve rendering verified
Earlier aggregate scores were withdrawn. Public weights await promotion.
Qwen on our compute
Deployment recipes for upstream models, with the configuration and measurement conditions recorded alongside the commands.
Qwen3.8 27B
Full multimodal stack, co-residency notes, and corrected token counts.
Qwen3.8 Flash-Next
A ~180B model with CPU / NVMe offload. Throughput remeasurement is pending.
Qwen3.6 35B-A3B
Recovered MoE configuration, downloadable launch script, and historical benchmarks.