Kimi K3.1 Pro, Nemotron Bolt, ThinkingCap, TB screening, LitPivot

Three open-weights releases lead today's log, followed by a health deployment and a research tool. I read the primary pages and ran none of the models, so the figures are what each publisher reports.

Kimi K3.1 Pro arrives, weights to follow

Kimi describes K3.1 Pro as its most capable 3T-class open-weight model to date, with a 1-million-token context window. A Highspeed variant runs 2 to 3 times faster. Both are live now on Kimi.ai, the mobile apps, Kimi Code and the API under the IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed, with Zero Data Retention available for API customers.

The weights are the part to read carefully. The post says full weights are planned by Nov 4, 2026 and were not yet available when it went up. Until then this is a hosted model with an open-weight promise.

The reported numbers are mixed in an honest way. On Terminal-Bench 4.0, K3.1 Pro scores 65.7%, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. On BenchCAD with tools it scores 92.7%, behind Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%. Kimi also reports claim-level hallucination on its Hard bench falling from 23.4% for K3 to 8.3%. For the open release, cyber-related data was removed from training, and the company says its cyber capability stays behind frontier proprietary models.

NVIDIA Nemotron 3.5 Bolt, small and local

NVIDIA post-trained Gemma-4-E4B into Bolt, built for low memory and fast time to first token. It is 8B total and 4B effective parameters, shipped as a 6.18 GB Q4_K_M GGUF, with 64K input and output. The license is OpenMDW-1.1, and NVIDIA notes the underlying Gemma 4 is Apache 2.0.

NVIDIA's own table is candid. Bolt beats the base Gemma on MultiChallenge (55.60 vs 39.36) and Tau2-Bench (52.32 vs 38.72), but Qwen3.5 4B leads on BFCL-v3, Tau2-Bench and RULER at 64K. The intended use is local voice assistants and embedded interfaces.

ThinkingCap trims Gemma's thinking

BottleCap AI released ThinkingCap-Gemma-4-26B-A4B, a reduced-thinking version of Google's gemma-4-26B-A4B-it. Across 12 benchmarks the average thinking-token cut is 38.9%, ranging from 13.7% on IFBench to 69.1% on LongBench v2. Average accuracy moves from 75.18% to 75.03%. The biggest drop is AA-LCR at 4.00 points, and looping falls from 3.64% of responses to 0.76%. The page does not state a license, so check the Hugging Face pages before building on it.

Google Research brings TB screening to the front line

Google's chest X-ray model is now inside a CE-certified tool from Nexus AI that screens an image in as little as 45 seconds. The post cites about 40 facilities in six countries and more than 25,000 radiographs evaluated by over 90 clinicians. The open side is MedGemma, at 4B and 27B, plus MedSigLIP and chest X-ray embeddings. FDA clearance is still a pathway, not an achievement.

Ai2 LitPivot

LitPivot splits an idea into problem, contribution and evaluation plan, then offers critiques backed by excerpts from retrieved papers. In a lab study of 17 researchers, two blinded experts rated LitPivot proposals higher on literature use than a baseline. The page gives no exact ratings and no code release.

What I'd watch next

The Nov 4 Kimi weights, and whether independent runs land near Terminal-Bench 4.0 at 65.7%. I also want to see ThinkingCap's license and third-party checks on whether the accuracy holds.