Inkling 2 Mini Benchmarks: ~200 tok/s on a Single RTX 4090
Thinking Machines' newly released small model Inkling 2 Mini has sparked a wave of local testing: several independent bloggers have verified its speed and coding capabilities on consumer hardware, with the overall verdict being strong model performance but an uneven deployment experience.
Confirmed
- Speed: eesahe got Inkling 2 Mini running at roughly 190-220 tokens/s on a single RTX 4090 (about 117 out of the box) by applying a few small patches to vLLM (which doesn't officially support running this model on a 4090), using an FP8+NVFP4 tested recipe. loktar00 achieved 170 tok/s on 2x3090, calling it an excellent local model, and is scaling up to a 4x3090 setup with more results and statistics to follow.
- Coding quality: Ubunta tested it on a Minisforum mini PC (Ryzen AI 9 HX 370, Radeon 890M iGPU, 89GB RAM, Linux) with roughly 4-bit quantization (model around 17GB), concluding that Inkling 2 Mini's code quality is genuinely strong, narrowly beating Qwen 3.8 27B on a self-built high-difficulty programming test.
- Motion design capability: Blogger Kyle Hessling1 found that, fed only the model card and benchmark data, Inkling 2 mini autonomously produced a motion graphic about itself in just 2 conversation turns—an area where small models usually struggle.
Not Yet Confirmed
- loktar00's 4x3090 scaling results and detailed statistics have not been published yet.
- Details of eesahe's vLLM patches and the full recipe await further public verification.
Why It Matters
- Inkling 2 Mini demonstrates that small models can combine high throughput with strong coding capability on single/dual consumer GPUs, offering direct reference value for the local deployment community.
- However, deployment barriers remain: vLLM doesn't officially support running this model on a 4090, and Ubunta explicitly said the PC side isn't worth the hype yet—it's a hassle to set up, though Mac users might give it a try. The gap between capability and usability is the current main point of contention.
2026-11-21 ~ 2026-11-21 · 5 related posts
Primary sources
- [source] Inkling 2 Mini hits ~200 tok/s on a single RTX 4090 with FP8 + NVFP4 quantization — eesahe · 2026-11-21
- Inkling 2 mini stuns in early tests: autonomous motion design in 2 turns at 104tps on RTX 5090 — LiTianleli · 2026-11-21
- [source] Inkling-2-Mini hits 170 tok/s on 2x3090s in local demos — LiTianleli · 2026-11-21
- [source] Inkling 2 Mini on PC: edges out Qwen 3.8 27B on coding tests but painful to run — Ubunta · 2026-11-21
- Inkling 2 Mini edges Qwen 3.8 27B on coding (277 vs 272) but struggles on PC hardware — Ubunta · 2026-11-21