Post-training for Ricursive Intelligence's Ripple-65B Preview incorporated synthetic data pipelines, long-context stages at 2 million sequence length, and reinforcement learning using truncated importance sampling

Overview Debates Related claims Topics Sources

Nous Research released Arete-27B, an open-weight supervised fine-tune of Alibaba's Qwen3.8-27B designed to remove Chinese state framing from model responses

Nous Research trained Arete-27B using supervised fine-tuning on approximately 53 million tokens over two epochs without reinforcement learning or preference optimization

The training dataset for Arete-27B contains 34,878 prompts across seven value categories, with 29,236 accepted teacher answers forming the fine-tuning split

Based on the fine-tuning results of Arete-27B, Nous Research stated that the state alignment bias layer in Chinese language models is shallow and applied late in training

Thinking Machines pre-trained Inkling 2 Small on an NVIDIA GB300 NVL72 cluster containing 72 Blackwell Ultra GPUs per system in 68 hours, followed by nine days of reinforcement learning

Baidu pre-trained ERNIE 5.5 Flash Base on 33.5 trillion tokens using the Muon optimizer with a 128,000-token training context window

Ricursive Intelligence reported requiring roughly 70 times fewer reinforcement learning rollouts for Ripple-65B Preview than the 100 million rollouts reported for Reflection Beam

Ricursive Intelligence reached target reinforcement learning performance for Ripple-65B Preview using 1.4 million rollouts

OpenRouter's hosting endpoint for Ripple-65B Preview currently restricts context length to 16,384 tokens

Fal chief executive officer Burkay Gur stated that Fal does not plan to pre-train its own models in the short term, though he did not rule out doing so in the future

Industry researchers Chandan Singh and Yuchen Jin suggested on X that weight editing could enable automated model maintenance, debugging, and AI-directed self-improvement without conventional fine-tuning

Mistral Large 4 was trained on 3,800 Nvidia Grace Blackwell GPUs in European data centers following three weeks of reinforcement learning after Mistral Large 4's October 2026 preview

Mistral Large 4 utilizes a sparsely activated architecture containing 1.05 trillion total parameters and 52 billion active parameters, offering a 1 million token context capacity

External AI researchers disputed Cohere's interpretation of Qwen3.8-Flash-Next's attempted upload, arguing it was likely an unsanitized training artifact from practice environments rather than a deliberate backdoor or exfiltration attempt

Google's Gemini 4 Carbon achieved a 96.2% score on the GraphWalks long-context benchmark spanning 256K to 1M tokens, compared to 84.2% for Gemini 4 Argon and 71.8% for GPT-6 Astra

Liquid AI's LFM3-1.7B base model was pretrained on 22 trillion tokens, with the final 3.5 trillion tokens dedicated to agentic mid-training to extend its context window to 256,000 tokens

Processing a 128,000-token context with Liquid AI's LFM3-1.7B requires 1.6GB of KV cache working memory during generation, compared to 4.3GB for Qwen3.5-4B

On the RULER 128K context evaluation, Liquid AI's LFM3-1.7B scored 65.56, trailing Qwen3.5-9B's score of 87.11

Liquid AI's LFM3 post-training pipeline comprised four stages: supervised fine-tuning on 1.75 trillion tokens, mixed reinforcement learning, distillation of six specialist teachers into one model, and agentic reinforcement learning

Cadence Maestro models are specialized, post-trained AI models grounded in electronic design automation tools, domain knowledge, curated engineering data, and deterministic solvers