On 2026-11-05 Nous Research released Hermes Agent v0.23.0 (The Hearth). Your gateway can serve the full Desktop in a browser; CLI/TUI/Desktop/editors/bots/cron share one session authority; the plugin catalog grew from ~600 to 1,050+; plus conversation rewind, inline terminal images, and paid-MCP spend limits. Official: ~1,950 merged PRs and 666 contributors.
Zoom
Key Takeaways
Official GitHub release: Hermes Agent v0.23.0 (2026-11-05), codename The Hearth
Webapp: gateway serves full Desktop in a browser (off by default; webapp.enabled)
Unified sessions: one gateway owns chats so surfaces stop double-writing
Plugin catalog ~600→1,050+ with a Verified tier; paid MCP tools can declare cost + chat spend caps
Independent evaluator Artificial Analysis published Tencent Hunyuan Hy4 results on the Intelligence Index v5 on November 5 (08:28 UTC): a composite score of 35, tying Grok 4.7 (xhigh) and one point behind Claude Haiku 5.5 (max). Among China-based labs it ranks second only to Kimi K3.1 Pro (46). On agentic coding it scores 31% on FrontierCode 2.0 and 37% on Terminal-Bench 4.0; cost per Intelligence Index task is about $0.78, roughly one-tenth of same-score Grok 4.7.
Zoom
Key Takeaways
Intelligence Index v5 composite 35: ties Grok 4.7 (xhigh), one point behind Haiku 5.5 (max, 36); No.2 China-lab model behind Kimi K3.1 Pro (46)
FrontierCode 2.0 at 31% (leads China labs; DeepSeek V4.1 Flash 1017 max 28%, Kimi K3.1 Pro 25%); Terminal-Bench 4.0 at 37%
AA-Omniscience v1.1 accuracy 50% but hallucination rate 89% (Kimi K3.1 Pro hallucination 37%); HLE-Diamond 33%
About $0.78 per Intelligence Index task: ~10x cheaper than same-score Grok 4.7, ~13x cheaper than Qwen3.8 Max (34), nearly 2x Haiku 5.5
Pricing about $0.834/$2.501 per 1M input/output tokens, cache reads $0.042; 1M context; text and image input
Ricursive Intelligence released its first model, Ripple-65B Preview, on November 4 (18:24 UTC), open-sourcing the instruction-tuned model (BF16 and NVFP4) and a base model on Hugging Face under Apache 2.0. The official model card describes a 65B dense Transformer pretrained on 12T tokens with up to 1M context, and self-reports TerminalBench 2.1 80.5, SWE-bench Pro v2.0 Hard 79.1 and DeepSWE 1.1 44.3, ahead of much larger US open-weight models on agentic coding. It fits on local hardware such as the DGX Spark. CEO Anna Goldie says it cost about a tenth of the prior comparable state of the art.
Zoom
Key Takeaways
Ripple-65B Preview is a 65B dense model with a thin-and-deep shape (80 layers, d_model 5120), moving from an Olmo 3 baseline to a Gemma 3N-like architecture.
It was pretrained on 12T tokens at 8K sequence length, then went through mid-training, long-context training, SFT and mixed-domain RL, with up to 1M context.
Self-reported TerminalBench 2.1 is 80.5, above Reflection Beam 501B (80.1) and Muse Spark Open 816B (76.4), but SWE-bench Verified 76.5 and DeepSWE 1.1 44.3 show it does not lead everywhere.
Ricursive says the core team grew from one to five people and code compute efficiency improved 16x versus the Olmo 3 ladder; it reaches about 250 tok/s on 2x B200.
Hugging Face hosts BF16, NVFP4, base and DFlash2 speculative-decoding checkpoints, all under Apache 2.0.
Cognition released SWE-2 Flash on November 4 (19:55 UTC). Its official blog says the model is built on Thinking Machines' Inkling-Small (276B total, 12B active parameters) and, for the first time, adds a large mid-training stage before reinforcement learning, lifting FrontierCode 2.0 from 3.2% to 24.0% and Terminal-Bench 4.0 from 0.9% to 20.0%. Cognition says it beats every American open-weight model. It generates about 200 tokens per second at one-fifth the price and 3x the speed of SWE-2, and is free for Pro, Max and Teams users in Devin Desktop and CLI until the end of November. Weights are not public; a secure sharing program for partners and researchers is planned.
Zoom
Key Takeaways
SWE-2 Flash is built on Thinking Machines' Inkling-Small, with 276B total and 12B active parameters, about one-tenth the size of SWE-2 (2.8T-A104B).
Training adds a three-stage mid-training run on a few trillion tokens of scientific articles, blogs, code and other sources, followed by large-scale RL with higher-quality and more diverse tasks.
Official scores: FrontierCode 2.0 24.0 (SWE-2 scores 23.8), DeepSWE 1.1 60.1, Terminal-Bench 2.1 89.2, SWE-Bench Pro v2 Hard 89.9 and Terminal-Bench 4.0 20.0.
It runs at about 200 tokens per second versus about 60 for SWE-2, at one-fifth the price, and is free for Pro, Max and Teams users in Devin Desktop and CLI through November.
Citing cyber risk, the weights are not openly released; Cognition says details of a secure weight-access program will follow in the coming weeks.
On October 30, Microsoft's Mage team released Mage-VL-Cowork on Hugging Face under Apache-2.0: a 4B multimodal computer-use agent in which a single model decides at every step whether to act through screenshots and GUI clicks (computer_use) or through bash and Python in a terminal, with no external orchestrator. Microsoft reports a two-run average of 66.0% on OSWorld-Verified (361 tasks, 50-step budget), above Qwen3.6-27B at 63.9% but below the 9B SCALECUA-Qwen3.5-9B at 68.7%.
Zoom
Key Takeaways
4B parameters, built on Mage-VL (Mage-ViT vision encoder plus Qwen3-4B-Instruct-2507); about 9.48 GB of bf16 weights under Apache-2.0.
On OSWorld-Verified, two independent runs scored 65.33% and 66.75% for a 66.0% average; 87.5% on operating-system tasks and 62.7% on multi-app workflows.
In Microsoft's side-by-side, a LibreOffice Calc chart task took 39 seconds (6 steps) via Python in the terminal, versus 9:56 (39 steps) for GUI-only Qwen3.6-27B.
Online RL adds 10.16 points over hybrid SFT while cutting mean model calls by 13.0% and output tokens by 25.4%.
Training data came from Mage-CUA-Forge, which synthesized about 294K desktop and web tasks across 6,915 skills and cut per-task generation cost from $0.906 to $0.158.
On Nov 2 Meituan's LongCat team listed LongCat-2.5 in its API platform change log as the full release of September 25's LongCat-2.5-Preview. The official technical blog says it has 1.6 trillion total parameters with about 48 billion activated per token, builds on LongCat-2.0 with stronger native visual understanding and agentic coding, and was trained and served with W4A4 inference on domestic AI accelerators. Vendor-reported scores include 88.4 on Terminal Bench v2.1, 38.5 on Terminal Bench v4.0, 91.9 on BrowseComp and 83.1 on MCP Atlas. The API speaks both OpenAI and Anthropic formats with a 1M-token context and up to 128K output tokens, at a limited-time price of $0.30 per 1M uncached input tokens, $0.006 cached input and $1.20 output.
Zoom
Key Takeaways
Scale: a 1.6-trillion-parameter MoE with about 48B active per token, a 1M-token context and up to 128K output tokens (official blog, API docs)
Code agents (vendor-reported): Terminal Bench v2.1 88.4 (Claude Opus 5 89.1, GPT-5.6-Sol 88.8), Terminal Bench v4.0 38.5 (GPT-5.6-Sol 37.3, Claude Opus 5 53.9), DeepSWE v1.1 64.6
General agents (vendor-reported): BrowseComp 91.9, MCP Atlas 83.1, AutomationBench v1.0.6 51.8, HLE with tools 63.1; AutomationBench jumps from 15.8 on LongCat-2.0
Pricing: limited-time $0.30 per 1M uncached input tokens, $0.006 cached input and $1.20 output (pricing page)
Engineering: W4A4 inference raises prefill throughput 43.3% and decode throughput 42.77% over a W8A8 baseline; FSDP gives 86% higher per-device throughput than DDP in long-sequence SFT
On Oct 30, SenseTime released SenseNova-Loop-9B-Preview (Apache 2.0) on its official X account and Hugging Face. It runs layers 4 to 27 of a 32-layer decoder twice with the same weights, so each forward pass executes 56 layer computations while storing about 9B parameters, about 14.6B at inference, and it uses two-stream mHC (manifold-constrained hyper-connections) to manage the residual streams. It takes text and images, with training sequences up to 262,144 tokens. SenseTime charts show it beating Qwen3.5 and Qwen3.6 peers on BIRD-Dev, WorkBuddy Bench, and other agent and data tasks, while it still trails on Terminal Bench 2.1; the technical report is not out yet. The official launch post drew about 910 likes.
Zoom
Key Takeaways
Structure: 32 distinct decoder layers, with layers 4 to 27 looped twice on shared weights for 56 layer computations per forward pass; about 9B stored and about 14.6B inference parameters (model card)
Weak spots from the same vendor chart: Terminal Bench 2.1 at 40.45, below Qwen3.6-27B at 47.19; tau3-Banking at 18.56, slightly below Qwen3.5-27B at 19.59
Training context up to 262,144 tokens; hybrid attention uses three linear-attention layers plus one full-attention layer per four-layer group; MTP speedup is available, DSpark and DFlash2 are still training
Apache 2.0, requires trust_remote_code=True; official examples cover Transformers 5.18.0, a 4-GPU LightLLM setup, and an SGLang PR, and the technical report is still to come
HarnessPro (arXiv 2610.26770, submitted Oct 19 and revised Oct 20), from Tsinghua University, Z.AI (Zhipu) and collaborators, is a label-free way to improve an agent harness: specialized agents diagnose execution traces, rewrite instructions and runtime policies, score candidates with LLM-generated tests, and keep Lite, Medium and Heavy harnesses along the performance-cost frontier. Per the paper, with GPT-5.5 and Codex the Heavy mode reaches 80.8% on Terminal-Bench 2, 4.5 points above the prior RHO method, while cost per task drops from $1.09 to $0.72. The code is MIT-licensed and drew wider media coverage this week, including from Pandaily.
Zoom
Key Takeaways
Terminal-Bench 2 (GPT-5.5 with Codex, mean of 3 runs): Heavy reaches 80.8% at $0.72 per task, versus RHO at 76.3% and $1.09, a 33.9% cost cut (paper figures, arXiv)
Averaged over three benchmarks: +2.0 points pass rate and -26.1% cost per task versus RHO, across Terminal-Bench 2, SWE-Bench Pro and ProgramBench
Holds across models: with GLM-5.2 and Claude Code on Terminal-Bench 2, Heavy hits 81.4% at $0.62, versus vanilla at 74.6% and $0.92 and RHO at 78.0% and $1.00
Budget modes: Medium matches RHO at 76.3% on Terminal-Bench 2 for only $0.56 per task; Lite gets 72.9% at $0.43
Cheap to train: only 10 training tasks per benchmark and 3 to 4 evolution epochs, tested on 59, 100 and 100 held-out tasks; the three harness tiers are MIT-licensed on GitHub
LM Studio posted on X at 15:32 UTC on Oct 29 that users can now download a new oMLX engine inside LM Studio and run it with MLX models they already have. The official changelog lists it in Bionic 1.1.10, alongside speculative decoding for MLX models and an in-app terminal. oMLX is an open-source Apple Silicon inference server maintained by developer jundot (Apache 2.0, about 23,000 GitHub stars) built around continuous batching and a two-tier RAM and SSD KV cache. LM Studio has not published speed comparisons for the engine.
Zoom
Key Takeaways
LM Studio says the oMLX engine can now be downloaded inside the app and used with existing MLX models; the post has about 540 likes (post)
The Bionic 1.1.10 changelog (Oct 29) also lists speculative decoding for MLX models, a public in-app terminal and llama.cpp 2.66.0 extension packs (changelog)
oMLX is Apache 2.0 open source with about 23,000 GitHub stars and needs macOS 15+ on Apple Silicon M1 to M5 (GitHub)
Its core is a two-tier KV cache: hot blocks stay in RAM, overflow is written to SSD as safetensors, and matching prefixes are restored from disk even after a restart
LM Studio has published no speed comparison with its existing MLX engine; one user reported that restart-to-update failed after launch and LM Studio is investigating
AI coding agent company Factory announced Droid 1.0 on X and its blog at 22:31 UTC on Oct 29, calling it the largest update yet to its agent harness. The key change makes programmatic tool calling (PTC) the default: the model writes a JavaScript program that runs up to hundreds of tool calls in a sandbox in one model request. In Factorys own Terminal-Bench 4 runs, Droid scored 57.6% with GPT-6.1 Sol ($0.79 per passed task) and 65.2% with Claude Opus 5.5 ($4.73), ahead of Codex, Claude Code, Cursor, OpenCode and Devin on the same models.
Zoom
Key Takeaways
Programmatic tool calling is now the default: the model writes JavaScript that runs in a sandbox with no file, network or system access of its own, making up to hundreds of tool calls per request, with independent calls run concurrently
Factory example: moving and verifying 151 Linear issues took 312 tool calls; direct calls needed about 36 requests and 3.1M input tokens, a Script needed 2 requests and 61K tokens
Terminal-Bench 4 (Factory runs, Oct 23-28, three trials per task): with GPT-6.1 Sol, Droid 57.6% vs Codex 54.0%, OpenCode 51.0%, Devin 45.0%; with Opus 5.5, Droid 65.2% vs Cursor and OpenCode 62.1%, Devin 59.6%, Claude Code 57.6%
The first request shrank from 20,887 to 13,055 tokens (-37%) and up-front tool definitions went from 11 to 2; newly loaded MCP tools are appended to keep the prompt cache intact, and a typical production request reads 98% of input from cache
Available now as a macOS and Windows desktop app, a terminal UI and the Droid SDK (@factory/droid-sdk); the benchmarks are vendor-run with no third-party reproduction yet
On Oct 29 Mistral released an improved Mistral Large 4 (Le Chonk) with open weights on Hugging Face (plus an NVFP4 quantized checkpoint), served on its API as mistral-large-4. After about three more weeks of reinforcement learning since the early-October preview, Mistral reports 19.7% on FrontierCode 2.0, Terminal-Bench 4.0 up from 28% to 41%, and new low/medium/high reasoning modes. The license is a modified MIT that allows commercial use, but running a model-as-a-service business needs a separate commercial license.
Zoom
Key Takeaways
Specs: about 1T total parameters with 49B active per token (about 52B including embeddings and output layers), 128 routed experts with 4 active plus 1 shared expert, 1M context, text and image input
Per Mistral: 19.7% on Cognition's FrontierCode 2.0, ahead of DeepSeek V4.1 Flash at 19.3% and GLM-5.3 at 17.8%, first among open models
48% on the Artificial Analysis Coding Agent Index, roughly tied with GLM-5.3 while using 47% fewer tokens per task on average; Terminal-Bench 4.0 rose from 28% in preview to 41%
51% on the Artificial Analysis Cyber Index, and a 96.6% average refusal rate on malicious cyber requests, which Mistral says is the highest among open models
Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters; in Surge Code blind pairwise tests it ranked behind only Claude Opus 5.5
On Oct 28 independent evaluator Artificial Analysis released Intelligence Index v5, its biggest revision since v4 in January: it adds Cognition's FrontierCode 2.0 and Terminal-Bench-Science, swaps Humanity's Last Exam for HLE-Diamond, and now draws 60% of its weight from private or held-out test sets. Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra at 57 and GPT-6.1 Sol at 55; the top open-weights model, Kimi K3.1 Pro (weights not yet released), scores 46, widening the gap from 3 points in v4.3 to 12. It also launched Progress Index v1 for comparing models across years.
Zoom
Key Takeaways
Overall (vendor-reported by Artificial Analysis): Claude Opus 5.5 (max) 58, GPT-6 Astra (max) 57, GPT-6.1 Sol (max) 55; Anthropic and OpenAI hold the top five, with Gemini 4 Argon (high) next at 50
The open-weights gap grows from 3 to 12 points: Kimi K3.1 Pro scores 46, while released-weight leaders DeepSeek V4.1 Flash 1017 and GLM-5.3 (max) score 32, below Claude Haiku 5.5 (max) at 36
Cost per task varies more than 13x: among the top 10, GPT-6.1 Terra (max) is cheapest at $0.96 and Claude Fable 5.1 (max) priciest at $13.04; GPT-6.1 Sol trails Opus 5.5 by 4 points at about one eighth the cost ($1.27 vs $10.04)
Four categories at 25% each: Knowledge Work, Coding (FrontierCode 2.0 15%, Terminal-Bench 4.0 10%), Science and General; GDPval-AA v2.1, SciCode and AA-LCR were removed
Tasks are heavier: median output tokens per task rose from 35k to 59k, and effort matters more, with Opus 5.5 scoring 35 at low effort vs 58 at max for 10x the cost
On Oct 28 DiffuSpace, whose members come from the HKU NLP group behind Dream 7B, released its first model, Clover-35B-A3B-Preview: a mixture-of-experts diffusion language model initialized from Qwen3.6-35B-A3B, with 35B total and about 3B active parameters, a 128K context window and block-parallel denoising. Vendor-reported results include 73.6 on SWE-bench Verified, 51.7 on Terminal-Bench 2.1 and 83.8 on GPQA Diamond, with 680 tokens/s on a single GB300 in agent workloads. Weights are Apache 2.0 on Hugging Face, ModelScope and GitHub, with SGLang and llama.cpp support.
Zoom
Key Takeaways
A 35B-total, roughly 3B-active MoE diffusion model initialized from Qwen3.6-35B-A3B, with 128K context and a thinking mode
DiffuSpace says it tops all six agent benchmarks it compared: SWE-bench Verified 73.6, Terminal-Bench 2.1 51.7 and tau3 Overall 65.3, ahead of LLaDA2.2 flash, Mercury 2.5 and Qwen3.6-35B-A3B
Speed (vendor-reported): 680 tokens/s on one GB300 at BF16 and batch size 1 in agent workloads, and 515 tokens/s on one H200 versus 365 tokens/s for Qwen with DFlash2
An INT4 build runs on a 16GB RTX 5070 Ti at 101.8 tokens/s on MATH-500
Weights on Hugging Face are Apache 2.0, about 69GB in BF16, and SGLang is the recommended serving path
Shanghai AI Laboratory has moved Atria Dawn from Preview to a full release, with open weights (plus an FP8 build) on Hugging Face under internlm/Atria-Dawn and on ModelScope, MIT licensed. Built on the 744B-parameter MoE GLM-5.2, it targets long-horizon research and engineering agent work; context doubles from the Preview's 256K to 512K and input is text only. The vendor table shows it leading the listed models on AutomationBench, CyberGym and SkillsBench, while still trailing Claude Opus 5 on coding tests such as SWE-bench Pro and Terminal-Bench 2.1 (all vendor-reported).
Zoom
Key Takeaways
512K context, double the Preview's 256K; built on the 744B-parameter MoE GLM-5.2, text input only
On Oct 26 OpenAI released GPT-6.1 Terra, an efficient model for visual reasoning, precise computer use and natural writing. It is rolling out today to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.1-terra. Standard API pricing is $1 input, $0.05 cached input and $5 output per million tokens, which OpenAI says is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol, with a 1,050,000-token context window. By OpenAI's figures it scores 15 points higher than GPT-5.6 Terra on Agents' Last Exam at medium effort for about a third of the cost, and solves 50.3% of Terminal-Bench 4.0 at max effort.
Zoom
Key Takeaways
Pricing (official docs): $1 input, $0.05 cached input, $1.25 for 30-minute cache writes and $5 output per million tokens; requests over 272K input tokens are billed at 2x input and 1.5x output; Batch/Flex are half price and Fast mode is 2x
Specs: 1,050,000-token context, 922,000 max input and 128,000 max output tokens, Jun 1 2026 knowledge cutoff; reasoning.effort from none to max, and tool calling requires the Responses API
Computer use (OpenAI figures): on Agents' Last Exam at medium effort it beats GPT-5.6 Terra by 15 points at about one-third the cost, and matches GPT-6 Astra at one-tenth the price per token; it finishes tasks over 3x faster than GPT-6 Astra and GPT-6.1 Sol
Coding and factuality (OpenAI figures): 50.3% on Terminal-Bench 4.0 at max effort, more than double GPT-5.6 Terra's published 21.5%; error rate on an internal error-inducing prompt set fell from 53.6% to 16.7% at low effort
Safety: OpenAI says it meets the Critical cybersecurity threshold under its Preparedness Framework and keeps the safeguards built for GPT-6 Astra
