The Hub for AI Coding Models, Tools & Workflows

On **3 Oct 2026**, Aleph Alpha released open-weight **Kolibri-1** ([HF](https://huggingface.co/Aleph-Alpha/Kolibri-1), **Apache-2.0**): 78B MoE / ~3.46B active, bilingual DE/EN, 262k native context (1M extrapolated), reasoning effort + tools. Official card: LiveCodeBench v6 **85.9**, SWE-Bench Verified **66.4**, TerminalBench 2.1 **27.7**. Serve via `aleph-alpha-inference` + vLLM.

Open-source [Louis-CFM/coucou](https://github.com/Louis-CFM/coucou) (MIT, ★3146) parks Claude Code/Codex/Cursor/Gemini CLI/Antigravity sessions in the Mac notch (top-edge island on Windows/Linux), with in-island Allow/Deny/Always approvals, live diffs, file drop, and service pills. macOS **v0.1.3** shipped 2026-10-03; Linux **0.1.1 beta** is out; Windows installer is paused for a Defender false positive.

Large language models are rapidly deployed across cybersecurity workflows to translate analysts' intent into command-line interface (CLI) executions. However, existing benchmarks emphasize either high-level knowledge QA or unconstrained agent rollouts, failing to measure precise parameter binding across real-world security tooling where minor syntax glitches invalidate execution. Researchers from MBZUAI introduce KaliBench, a fine-grained natural-language-to-CLI benchmark on Kali Linux spanning 8,504 query-command pairs across 1,642 tools, 23 capability dimensions, and 5 security phases. Auditing reveals no open-weight model surpasses 42% exact accuracy without tool hints. Furthermore, reinforcement learning with KaliBench's runtime-free verifiable rewards elevates an 8B model to rival a 685B MoE frontier model.

Conventional LLM agents emit reasoning thoughts and actions token-by-token, making extended interactions slow and compute-heavy. While Jev-style probabilistic models deliver rapid predictions over finite action domains, they mandate pre-specified, static fields, precluding autonomous adaptation in open-ended language tasks. Shanghai Jiao Tong University researchers introduce JevSpawn, a compositional policy framework bridging natural language instructions with finite probabilistic exploration. JevSpawn couples parallel action spawning with feedback-driven branch selection, representation revision, and recovery from cached alternatives. By sharing prefix KV caches and action structures, JevSpawn eliminates repeated context generation without model training, systematically outperforming seven leading agent baselines across eight benchmarks.

On-policy self-distillation (OPSD) trains mathematical reasoning models by deploying a privileged teacher that observes ground-truth solutions to supervise student-sampled prefixes. Standard OPSD fixes the teacher parameters across all states, leaving valuable supervision untapped. Researchers from Tencent WeChat and Tencent AI Lab propose Neighborhood OPSD (N-OPSD). They discover that local parameter perturbations in the privileged teacher reveal complementary, reference-aligned corrections across distinct token positions. Offline greedy pruning compiles a compact pool of frozen neighbor experts, while an online MaxPeak and quantile router dynamically selects the optimal expert distribution. Across AIME 2024, AIME 2025, and HMMT 2025, N-OPSD improves Average@12 by up to 2.75 points on Qwen3-1.7B, 4B, and 8B models without any inference overhead.

Official Antigravity Models docs now list Claude Sonnet 5.5 (thinking) and Claude Opus 5.5 (thinking): Free/Plus/Enterprise ❌; Google AI Pro non-trial only ✅; Ultra ✅. Footnote: Claude 4.6 and GPT-OSS-120b remove on 2026-11-02. No separate changelog post—docs table is the source. No invented Antigravity SWE scores.

At DevDay (2026-09-29) OpenAI shipped reusable Codex Cloud environments: publish a project setup (repos, deps, tools, network secrets), then each task runs in its own VM workspace and can continue while your laptop sleeps. Docs: learn.chatgpt.com/docs/cloud and Help Center. Rolling out to Plus/Pro and workspace plans; saved VM state recoverable ~7 days by default. No invented SWE scores.

stablyai/orca positions itself as an Agent Development Environment (ADE): parallel CLI coding agents (Claude Code, Codex, OpenCode, Pi, …) each in an isolated git worktree, with desktop, mobile companion, and remote hosts. Verified now: ★84,136 / 5,424 forks; MIT; desktop v1.4.219 (2026-10-02); Android mobile-android-v0.0.52 (2026-10-03). Site onOrca.dev. No official SWE leaderboard—no invented scores.

Before invoking tools, agentic LLMs navigate a K-way action space: executing calls, clarifying ambiguities, answering directly, or declining. While internal activation steering attempts to align pre-execution tool decisions, aggregate metrics obscure where manipulated latent states land and the severe collateral damage they inflict. Researchers from Aberdeen and Oxford introduce SAKIKO, an auditing framework formalizing representation repair via directional error discovery, router-conditioned intervention, destination-resolved verification, and prospectively frozen statistical licensing. Crucially, destination auditing demonstrates that behavioral movement does not equal repair: an intervention scoring a net +55 gain corrupted more than half of baseline-correct decisions, establishing the necessity of outcome-resolved adjudication.

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.

Real-world embodied tasks require agents to operate physical tools under geometric constraints while tracking object state transitions. Despite strong perception benchmarks, current multimodal video LLMs struggle with tool-centric physical reasoning. NTU S-Lab and collaborators introduce EgoTools, the first comprehensive suite for egocentric tool-use reasoning. It features EgoTools-Data (100 hours of synchronized egocentric tool videos with 3D annotations, audio, and dense causal narrations) and EgoTools-Bench (1,000 QA pairs across perception, geometry, procedural progress, and causal inference). Probing reveals frontier models struggle with physical grounding (Gemini-3.1-Pro scores only 51.7% on grounding), while supervised fine-tuning on EgoTools-Data lifts Qwen3-VL-8B-Instruct accuracy from 50.0% to 60.9%.

EverMind-AI/Raven orchestrates built-in Research/Code/Design/Oncall plus third-party harnesses (Claude Code, Codex, Hermes, Pi, …) via a Host Agent DAG with admission checks, and frames controlled self-evolution of Raven’s own harness (RSI). Verified: ★5097 / 145 forks; v0.2.3 (2026-09-27); arXiv 2609.33439. Site: Raven-Research 76.5% DeepResearch Mixed; paper: MAOB Exact Match +10.4/+10.5 pp vs strongest baseline. Pre-alpha.

Coding agents frequently reproduce existing code, error logs, and multi-turn refactoring traces, making retrieval-based speculative decoding (SD) an ideal acceleration paradigm. However, conventional retrieval drafters fail in agent workflows: candidate code on disk diverges from the agent's emission formats (e.g., diff blocks, escape characters), and static draft lengths ignore inter-agent turn drift. KAIST researchers introduce AgSpec, a speculative decoding framework tailored for coding agents. AgSpec constructs format-aligned corpora across session trajectories, workspaces, and global repositories, coupling offline draft caps with online feedback-adaptive draft length control. Across repository-level multi-agent benchmarks, AgSpec outperforms five retrieval drafters and EAGLE-3, boosting generation throughput up to 4.76x.

Vision-language model (VLM) agents control robots via visual feedback and motion primitives, but repetitive model queries and redundant visual frames inflict massive token overhead. Peking University researchers introduce PyRUA-Lean, an interactive code-execution framework coupling feedback-driven primitive composition with selective observation. Rather than emitting atomic tool calls, the agent composes robot primitives and learned vision-language-action (VLA) policies into executable Python cells that execute localized conditional checks and retries, requesting sensory images only upon necessary replanning. Across 700 tasks on LIBERO-PRO, RoboTwin 2.0, and RoboCasa365, PyRUA-Lean boosts success from 63.1% to 71.7% under identical budgets, slashing LLM calls by 49% and input tokens by 65%.

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.

OpenRig wraps multiple coding harnesses into YAML-defined rigs with seats, queues, and a local daemon/TUI so Claude Code, Codex, and supervised Pi run as one recoverable team. Verified now: GitHub ★4482 / 306 forks; npm @openrig/cli 0.6.4 (2026-10-02); ~4560 weekly downloads.

CodeAF is an Apache-2.0 Go binary that runs open-weight models as a local software factory with task fan-out and Pareto crewing. v0.6.0 shipped 2026-10-01; ★259 now. Vendor DeepSWE doc: /senior-dev solved 62/113 (54.9%) on DeepSeek V4 Flash vs nine other harnesses—with stated one-seed and sampling caveats.

Coding agents are moving beyond virtual codebases into the physical world of robotics. However, existing benchmarks primarily focus on isolated control policies, overlooking the broader engineering capabilities required of real-world robotics engineers. Researchers from Harvard, MIT, and Stanford introduce RLE-Bench, a comprehensive benchmark and qualifying exam for coding agents as Robot Learning Engineers. Spanning interactive control, policy learning, perception/estimation, and mechanical design, RLE-Bench evaluates agents across heterogeneous artifact synthesis, multimodal feedback reasoning, and hardware-constrained execution, introducing the aggregate RLE Index to benchmark frontier agent engineering capabilities.

Looped Transformers achieve dramatic parameter efficiency by repeatedly executing a single shared block across recurrent cycles. However, standard autoregressive decoding discards all intermediate recurrent states, utilizing only the final pass. Researchers from UT Austin, Princeton, and collaborators introduce LoopCD, a training-free contrastive decoding framework. By contrasting final recurrent predictions with early-stage weaker passes in logit space (LoopCD-Logits) or hidden space with zero output pass overhead (LoopCD-Hidden), LoopCD raises Ouro-2.6B-Thinking's AIME 2024 Pass@1 from 61.88% to 73.33% and Huginn's HumanEval Pass@1 from 22.56% to 31.71%. Crucially, LoopCD allows halving recurrent loops while outperforming full-depth unguided baselines, reducing forward FLOPs by 22.5% to 48.2%.

Streaming video LLMs must preserve perceptual evidence before downstream intent is revealed and respond proactively when evidence matures. Nanjing University researchers introduce OneStreamer, a unified architecture that couples query-independent factual recording with task execution through a shared proactive generation interface. Its Proactive Hierarchical Caption Memory (PHCM) synthesizes time-anchored micro-captions and event summaries, allowing inference to complement recent visual frames with cached factual records without reprocessing raw historical visual features. Trained on the broad-coverage OneStreamer-1M dataset, the 4B model establishes top SOTA across all eight streaming video understanding benchmarks.