# Large language models

AirLLM pushes large-model inference memory down to a few gigabytes on a single card: Kimi K3 (2.8T) measured end to end at 3.72GB of VRAM on one RTX 6000 Ada, Qwen3.8-Flash-Next (125B) at 5.95GB on one RTX 4090, and DeepSeek-V3 (671B) at about 12GB. It uses no quantization, distillation or pruning; instead it streams weights layer by layer, and for MoE models loads only the experts each token actually routes to. The project extends the same idea to training: frozen weights stream in one layer at a time while adapters stay on the GPU, letting a 125B model train under 6GB.

AirLLM pushes large-model inference memory down to a few gigabytes on a single card: Kimi K3 (2.8T) measured end to end at 3.72GB of VRAM on one RTX 6000 Ada, Qwen3.8-Flash-Next (125B) at 5.95GB on one RTX 4090, and DeepSeek-V3 (671B) at about 12GB. It uses no quantization, distillation or pruning; instead it streams weights layer by layer, and for MoE models loads only the experts each token actually routes to. The project extends the same idea to training: frozen weights stream in one layer at a time while adapters stay on the GPU, letting a 125B model train under 6GB.

Clinical prediction models often arrive as files without the data they were trained on, so an auditor must work out what the model actually computes. Researchers trained tanh networks of 105 to 38,401 weights on simulated subarachnoid haemorrhage cohorts, planted known relationships or a label bias, and asked GPT-6 Astra to answer with no tools, with a calculation-only white-box tool, or with a black-box tool exposing only predictions. Without tools Astra recovered planted relationships in 353-weight networks but became unreliable at 3,457 weights; with the calculator it recovered the relationship in all 29 runs that had one, raised no false alarm in 15 null runs, and black-box matched white-box.

Clinical prediction models often arrive as files without the data they were trained on, so an auditor must work out what the model actually computes. Researchers trained tanh networks of 105 to 38,401 weights on simulated subarachnoid haemorrhage cohorts, planted known relationships or a label bias, and asked GPT-6 Astra to answer with no tools, with a calculation-only white-box tool, or with a black-box tool exposing only predictions. Without tools Astra recovered planted relationships in 353-weight networks but became unreliable at 3,457 weights; with the calculator it recovered the relationship in all 29 runs that had one, raised no false alarm in 15 null runs, and black-box matched white-box.

Preferred Networks released PLaMo 3 Translate, a translation-specialized LLM suite originally built for its own PLaMo Translate service, as open weights on Hugging Face in four sizes (0.8B, 2B, 8B, 31B), runnable entirely offline on smartphones, laptops or internal corporate servers across 53 languages including English and Japanese. Under the PLaMo Community License, individuals and enterprises with annual revenue of 1 billion yen or less in their most recent fiscal year can download and use the models free for commercial and non-commercial purposes; larger organizations need a commercial license. PFN says its comparative evaluations put even the smallest 0.8B version ahead of comparable translation-specific LLMs such as Google's TranslateGemma and Tencent's Hy-MT2.

Preferred Networks released PLaMo 3 Translate, a translation-specialized LLM suite originally built for its own PLaMo Translate service, as open weights on Hugging Face in four sizes (0.8B, 2B, 8B, 31B), runnable entirely offline on smartphones, laptops or internal corporate servers across 53 languages including English and Japanese. Under the PLaMo Community License, individuals and enterprises with annual revenue of 1 billion yen or less in their most recent fiscal year can download and use the models free for commercial and non-commercial purposes; larger organizations need a commercial license. PFN says its comparative evaluations put even the smallest 0.8B version ahead of comparable translation-specific LLMs such as Google's TranslateGemma and Tencent's Hy-MT2.

Before the European Society of Cardiology Congress 2026 opened, three large language models locked predictions of whether each trial's primary endpoint would be met, using only design information public at registration. Of 48 scorable trials, 28 were positive; Claude, GPT and Gemini separated positive from negative trials with AUCs of 0.80, 0.79 and 0.74, all above chance. Among 26 trials whose effect was reported exactly as locked, the direction was correct in 88% of predictions, and the median multiplicative error of ratio estimates was about 15%.

Before the European Society of Cardiology Congress 2026 opened, three large language models locked predictions of whether each trial's primary endpoint would be met, using only design information public at registration. Of 48 scorable trials, 28 were positive; Claude, GPT and Gemini separated positive from negative trials with AUCs of 0.80, 0.79 and 0.74, all above chance. Among 26 trials whose effect was reported exactly as locked, the direction was correct in 88% of predictions, and the median multiplicative error of ratio estimates was about 15%.

Large language models write stories that often lack the rising and falling reader experience that makes narrative engaging. A new framework first learns how reader responses such as suspense or humor evolve in high-quality human-written stories, then uses a small model to predict the response the next segment should elicit and steers an off-the-shelf writer model with that prediction. Across suspense and humor, with GPT-5.5 and Qwen3.5-27B as writers, generated stories improve trajectory-based fidelity to human-written stories by 12%–35% while largely preserving story coherence.

Large language models write stories that often lack the rising and falling reader experience that makes narrative engaging. A new framework first learns how reader responses such as suspense or humor evolve in high-quality human-written stories, then uses a small model to predict the response the next segment should elicit and steers an off-the-shelf writer model with that prediction. Across suspense and humor, with GPT-5.5 and Qwen3.5-27B as writers, generated stories improve trajectory-based fidelity to human-written stories by 12%–35% while largely preserving story coherence.

Autoregressive language models normally learn next-token prediction in a fixed left-to-right order; Max(MI(S;t)) instead ranks target tokens by their mutual information with the source and already-selected targets before training, exposing informative targets earlier. On GPT-style models it reaches 94.96% average accuracy across arithmetic tasks, against 74.84% for next-token prediction and 86.29% for reverse order, close to the 97.33% upper bound from exhaustively testing 24 orders; on WikiText-2 it lowers average perplexity from 13.58 to 10.77. Only the training-target organization changes; architecture and autoregressive decoding stay the same.

Autoregressive language models normally learn next-token prediction in a fixed left-to-right order; Max(MI(S;t)) instead ranks target tokens by their mutual information with the source and already-selected targets before training, exposing informative targets earlier. On GPT-style models it reaches 94.96% average accuracy across arithmetic tasks, against 74.84% for next-token prediction and 86.29% for reverse order, close to the 97.33% upper bound from exhaustively testing 24 orders; on WikiText-2 it lowers average perplexity from 13.58 to 10.77. Only the training-target organization changes; architecture and autoregressive decoding stay the same.