AI Coding Live News

At its first Forge conference on November 3, open-model inference platform Fireworks launched a batch of products. Ultra-fast, powered by NVIDIA Blackwell, is in public preview at up to 1,500 tokens per second on DeepSeek V4.1 Flash, cutting a 30-turn agent task from 1 minute to 8 seconds. Background Mode runs deferrable work at half price, alongside multi-day prompt caching and higher rate limits. Nexus Desktop brings open-model routing into the Codex app, Claude Desktop and OpenCode, with Pro plans from $8 per month. Fireworks also announced No Eval, serverless LoRA inference, the Fireworks Sovereign private preview and a 60+ partner Frontier Ecosystem.

Zoom

On Oct 26, Entire co-founder and CEO Thomas Dohmke announced on X and the company blog that Entire is live as "the Git forge for agent-human collaboration." It has three parts: Trails, which replace diff-based review with eval scores computed on every push; Git-compatible distributed hosting that can replicate across regions or mirror from GitHub; and Continuous Memory, which stores agent sessions in the repo. Buildkite and Depot are the first CI integrations, and it is open to any developer today.

Zoom

At GTC Berlin, Mistral AI and NVIDIA announced a strategic partnership dedicated to advancing AI for industrial engineering, computational physics, and electronic design automation (EDA). The alliance couples Mistral frontier architectures with NVIDIA CUDA-X libraries, Warp physics simulation frameworks, and Omniverse digital twins. Mistral simultaneously revealed architectural specifics for its forthcoming flagship, Mistral Large 4: spanning roughly 1 trillion total parameters with 49 billion active parameters (MoE), trained on 3,800 Grace Blackwell GPUs, with open weights slated for official release at the end of October 2026.

Zoom

Researchers introduced MemPilot (arXiv:2610.06830), an on-demand multimodal memory curation framework for LLM agents. Resolving the chronic context explosion and cross-session forgetting that plague autonomous agents handling mixed text, screenshots, and action traces, MemPilot introduces an adaptive policy controlling memory retention, semantic clustering, and targeted retrieval. Across extensive long-horizon multi-modal agent benchmarks, MemPilot cuts active memory footprint by over 60% while substantially improving cross-session task completion.

Zoom

As AI agents transition from narrow tasks to long-horizon, cross-domain workflows, manual harness engineering has hit an insurmountable complexity ceiling: brittle bespoke scaffolding resists scaling, while domain coupling limits generality. EverMind AI introduces Raven, 'The Harness of Harnesses,' an open-source multi-agent ecosystem that autonomously synthesizes, evolves, and orchestrates modular harnesses tailored to distinct models and problem spaces. Raven encapsulates executable model-harness pairs as composable units of intelligence. Led by a Host Agent managing decomposition, agent matching, and result synthesis, and backed by EverOS and Skill Forge for persistent experience accumulation, Raven proves mathematically and empirically that dynamic harness composition reliably expands task coverage beyond monolithic baselines under matched resource budgets.

Zoom

The thesis that on-policy distillation inherently outperforms offline supervised fine-tuning by reducing catastrophic forgetting and enhancing generalization has become an accepted pillar of post-training. However, researchers from the University of Cambridge led by Mihaela van der Schaar present a definitive study (arXiv:2609.35259, 160+ upvotes on Hugging Face) disentangling rollout policies, token-level KL divergence directions, and optimization dynamics across Llama-3 and Qwen-2.5 models. The empirical findings reveal that rollout policy does not play the primary role: Forward KL is remarkably invariant and resilient across student and teacher rollouts, while Reverse KL is hyper-sensitive and strictly necessitates student-sampled trajectories. Furthermore, catastrophic forgetting and parameter update sparsity are governed predominantly by learning rate schedules, reframing the design space for LLM distillation.

Zoom

Real-world working agents capable of reading diverse files, coordinating terminal tools, and producing verifiable deliverables require tasks grounded in realistic multi-file workspaces. Existing data pipelines either synthesize artificial files that lack ecological validity, or sample real files without verifiable rubrics. Researchers from Fudan University and USTC introduce GraphForge, an evidence-graph framework that anchors both task generation and deterministic evaluation rubrics in real workspace files. Fine-tuning Qwen3.6-27B on only 2,169 GraphForge trajectories catapults GDPVal to 1445.7 (+65.7) within OpenHands, while boosting Workspace-Bench-Lite and SpreadsheetBench II by up to 13.7 points under Claude Code.

Zoom

Expanding text embedding models to multimodal domains typically triggers catastrophic forgetting in text retrieval, with existing omni-modal embedders ballooning past multi-billion parameters to compensate. MBZUAI researchers introduce Omni-Embed-Mini, a compact 0.9B architecture mapping text, speech, audio, image, video, and rich documents into a single shared cosine space. Crucially, text-side weights remain bit-identical to the base model, guaranteeing zero regression on text retrieval (49.57 nDCG@10 on MTEB-v2 BEIR-8). Powered by dense cascaded caption distillation, Matryoshka SigLIP loss, and online hybrid hard-negative mining, Omni-Embed-Mini is 2.7x to 9.5x smaller than competing open models, while its 2.3B variant surpasses proprietary gemini-embedding-2 on cross-modality averages.

Zoom

On-policy learning is widely assumed to mitigate catastrophic forgetting, induce sparser parameter updates, and enhance generalization in LLM post-training. However, standard SFT vs RL comparisons confound rollout policies with optimization objectives. Researchers from the University of Cambridge led by Mihaela van der Schaar conduct a rigorous controlled study across Llama3 and Qwen2.5 on reasoning, scientific, and medical tasks. Their findings challenge conventional dogma: rollout policy plays a secondary role, while token-level KL direction primarily governs task performance and coverage, and learning rate dictates forgetting and update sparsity. Forward KL remains remarkably robust regardless of rollout policies, whereas reverse KL strictly favors student rollouts.

Zoom

Hono creator @yusukebe amplified Cloudflare’s launch of cf (npm i -g cf): Forge generates 3000+ CLI ops from OpenAPI (vs Wrangler’s ~280), JSON-default output, cf cli search, TypeScript cloudflare.config.ts, and Vite-default local dev. Open beta [email protected].

Zoom