VQV Signal

OPEN SOURCE WATCH TECHNICAL

Builders report Inkling 2 Mini running locally, with early speed and eval figures

According to a post on X by @soumithchintala, Inkling 2 Mini decodes at roughly 75 tokens per second on a DGX Spark, with evals said to be roughly on par with Qwen 3.8 27B. Other builders report about 60 tokens per second on an RTX 3090 at 4-bit, and say it runs on a Mac mini. Artificial Analysis puts it at 20 on its Intelligence Index, Ollama and Unsloth have shipped support, and ggml-org has posted GGUF quants, with the IQ4_NL file at 17.6 GB. The snippet also says @eliebakouch highlighted that the team pruned two-thirds of the experts in mid-training.

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-11-20T23:45:00+00:00 Detected 2026-11-21T09:03:01+00:00
View original source

According to a post on X by @soumithchintala, Inkling 2 Mini decodes at roughly 75 tokens per second on a DGX Spark, with evals said to be roughly on par with Qwen 3.8 27B. Other builders report about 60 tokens per second on an RTX 3090 at 4-bit, and say it runs on a Mac mini. Artificial Analysis puts it at 20 on its Intelligence Index, Ollama and Unsloth have shipped support, and ggml-org has posted GGUF quants, with the IQ4_NL file at 17.6 GB. The snippet also says @eliebakouch highlighted that the team pruned two-thirds of the experts in mid-training.

AI-assisted summary based on the listed source.

@soumithchintala reports roughly 75 tok/s decode on a DGX Spark and evals roughly on par with Qwen 3.8 27B; others report about 60 tok/s on an RTX 3090 at 4-bit and runs on a Mac mini. Artificial Analysis puts it at 20 on its Intelligence Index, Ollama and Unsloth shipped support, and ggml-org posted GGUF quants...

Developers who want to run open-weight models on their own hardware now have early local-run reports, quantized files and tooling support to test. These figures come from social posts and third-party measurements, so results on other setups may differ. Watch for independent evaluations and whether the reported speeds hold on more consumer hardware.

Signal Strength 95% Technical label WATCH Public Interest 40 Category OPEN SOURCE Reader Depth TECHNICAL

Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.

Public Interest components
Recognizable Entity Score 60 Practical Impact Score 8 Novelty Interest Score 70 Consequence Score 18 Curiosity Score 16 Shareability Score 55

VQV surfaced this signal because it is recent, relevant to Open Source LLMs, connected to Scobleizer News Pulse (X).