AI Tools Weekly — Issue #2

AI Tools Weekly — Issue #2 | Oct 14 – Oct 21, 2026

> This issue's focus: new AI tools and reviews. Every item was checked against official pages (official site / official blog / GitHub Release / Hugging Face), verified on 2026-10-21 (Beijing time). Dates are those shown on the official pages. Our commentary is based on official materials, not hands-on testing by StackLoom. Benchmark results are vendor-reported unless stated otherwise.

1. Mercury 3 (Inception)

- Official link: https://www.inceptionlabs.ai/blog/introducing-mercury-3 - Release date: 2026-10-19 - Key features: A diffusion language model, billed as "the most capable diffusion LLM to date"; 1,068 tokens/s on NVIDIA GPUs; 1M-token context window (Mercury 2.5: 260K); more than 40x the parameters of Mercury 1; OpenAI API compatible. Available via the Inception API platform, OpenRouter, Baseten and Vercel. - Our take: Inception says quality matches or beats GPT-6 Luna and Claude Haiku 5.5 at 4–8x the speed. According to SiliconANGLE, it wins 8 of 12 benchmarks but trails Haiku 5.5 on DeepSWE (46.3 vs 51.6), and there is no independent third-party testing yet. A fit for latency-sensitive uses such as browser agents, real-time tutoring, search and customer support. - Pricing: $0.20 input / $0.70 output per million tokens; half price during launch ($0.10 / $0.35); cached input a further 90% off the input price; new API keys get 100 million free tokens. - Sources: official link; https://siliconangle.com/2026/10/20/inception-says-mercury-3-beats-gpt-6-luna-and-claude-haiku-5-5-at-up-to-8x-the-speed/

2. MiniMax M3.1 Flash (MiniMax)

- Official link: https://www.minimax.io/blog/minimax-m31f - Release date: 2026-10-19 - Key features: A new-generation Flash model focused on agentic tasks, coding, Computer Use and cybersecurity; MiniMax says online inference is nearly 3x faster and usage cost is 70% lower. "Space Bunny", released anonymously earlier, was an early version of it. Available via the MiniMax API, M Plan and MiniMax Code. - Our take: Self-reported scores posted on MiniMax's official X account: DeepSWE 74%, Automation-Bench 62%, ALE 30%, BabyVision 91% (not reproduced by third parties). The official X account says weights will be released on Oct 23 (the official blog only says "later"). Unverified: the weights release date. A fit for low-cost agent and coding workloads with heavy call volume. - Pricing: $0.15 input, $0.60 output, $0.015 cached input per million tokens (flat rate) - Sources: official link; https://xitter.cf/MiniMax_AI/status/2112150867868541221 (mirror of the official X post); Reuters report via: https://www.investing.com/news/stock-market-news/chinas-minimax-launches-updated-ai-model-m31-flash-claims-major-improvements-4957342

3. Muse Spark Open (Meta)

- Official link: https://research.meta.ai/blog/muse-spark-open ; model card: https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B - Release date: 2026-10-15 (Fireworks' launch post is dated 10-20) - Key features: A natively multimodal MoE open-weight model, 816B total / 42B active parameters, 1M-token context; focused on coding and long-horizon agentic tasks; runs locally on two M5 Ultra Mac Studios. Available via Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together AI. - Our take: Third-party results cited by Fireworks: ranked No. 1 (25.4%) on Vals AI's Harvey legal agent benchmark; 33% hallucination rate on AA-Omniscience, lower than Claude Opus 5.5 (59%). Downsides: self-hosting is demanding (NVFP4 needs ≥700GB of GPU memory, BF16 needs 2TB); released under the Muse Model License, with restrictions on commercial use. - Pricing: Weights are free to download (subject to the license); API pricing varies by platform. Unverified: API prices were not checked for this issue. - Sources: official link; https://pytorch.org/blog/muse-spark-open/ ; https://fireworks.ai/blog/meta-muse-spark-open

4. Voxtral 3 (Mistral AI)

- Official link: https://mistral.ai/news/voxtral-3 - Release date: 2026-10-20 - Key features: A family of speech models plus a voice agent SDK (built on LiveKit). Voxtral Small/Mini Transcribe 3 support 50+ languages with speaker diarization and context biasing; Realtime Transcribe 3 has under 500ms latency; Voxtral Mini TTS 3 supports 35+ languages and 50+ dialects with emotion control. Use it via the Mistral API or self-host. - Our take: Mistral says transcription beats Whisper, Scribe v2 and Gemini 3.5 Transcribe (vendor-reported). It connects the full chain from listening, thinking and tool calling to speaking, which suits call centers and voice assistants; preset voices include Mandarin. - Pricing: Mini Transcribe 3 $0.0035/min; Realtime 3 $0.008/min; Small Transcribe 3 $0.0040/min; Mini TTS 3 $13 per million characters; voice agent API $0.04–0.08/min. - Sources: official link - Correction: The sub-model name "Voxtral TTS 3" listed in our first draft was inaccurate; the official table lists it as "Voxtral Mini TTS 3".

5. Amazon Polly Generative engine upgrade (AWS)

- Official link: https://aws.amazon.com/about-aws/whats-new/2026/10/polly-generative/ - Update date: 2026-10-20 - Key features: The generative engine moves to a new model, with more natural, accurate and lower-latency speech; 14 new voices across 9 regions; Hindi and Finnish supported on the Generative engine for the first time. - Our take: A fit for companies already on AWS that need multilingual TTS; no third-party audio-quality comparison yet. - Pricing: Generative voices $30 per million characters; free tier of 100,000 characters per month for the first 12 months (the pricing page doesn't say whether this upgrade changes prices). - Sources: official link; https://aws.amazon.com/polly/pricing/

6. QVAC TranscribePsy (Tether AI Research)

- Official link: https://qvac.tether.io/blog/on-device-speech-recognition-transcribepsy ; https://huggingface.co/blog/qvac/transcribepsy-250m - Release date: 2026-10-14 (the official blog text says it was released "14 October 2026"; the page header date is Oct 13) - Key features: A 254M-parameter on-device speech recognition model in 4 versions (laptop with 25 languages / 8 languages, mobile GGUF, iPhone Core ML). - Our take: Self-reported results: 8.02% WER on the Open ASR Leaderboard, better than Whisper-small (12.34%); transcribes 10 minutes of audio in just 1.65 seconds on an iPhone 17 Pro Max. Limits: European languages only (no Chinese); batch transcription of whole files only, no streaming. A fit for offline, privacy-focused mobile apps. - Pricing: Free and open source (Apache 2.0) - Sources: official link

7. Sunra MCP (Sunra)

- Official link: https://sunra.ai/blog/introducing-sunra-mcp ; https://sunra.ai/mcp - Release date: 2026-10-19 — ⚠️ Unverified: the official blog post isn't dated. This date comes from an official X post indexed by the third-party site launcharchive.ai; the post says its price-comparison data was collected Oct 14–17, which is consistent with that date. - Key features: A remote MCP server (https://mcp.sunra.ai, OAuth sign-in) for calling image, video, audio and 3D models from tools such as Claude, ChatGPT, Claude Code, Codex and Cursor. - Our take: For models whose vendors publish API prices, it bills at the vendor's list price with no markup. Sunra openly admits not every model is the cheapest and publishes its comparison table. A fit for creators who want to centralize multimedia generation inside their AI assistant. - Pricing: Credit-based, 1 credit = $0.01 (Basic plan $12/month includes 1,200 credits); Pro plans $29–99/month, Creator plans $125–2,051/month, cheaper when billed annually. - Sources: official link; https://launcharchive.ai/categories/ai/2026/october

8. Codex CLI 0.163.0 (OpenAI)

- Official link: https://github.com/openai/codex/releases (tag rust-v0.163.0) - Update date: 2026-10-15 (GitHub Release time) - Key features: New /config settings search (supported settings apply instantly); send follow-ups while a reply is in progress to steer it in real time; /keymap supports a leader key (default Ctrl+X); /daemon restarts the background service; read-only tools can run in parallel; Daybreak enabled by default for eligible accounts. - Our take: Mostly experience and stability improvements (plus fixes for permission issues after resuming a session, network proxy security issues and more); practical for heavy command-line users. 0.164.0 is still an alpha pre-release. - Pricing: The CLI is open source; model usage is billed to your OpenAI account. Unverified: model usage pricing was not checked for this issue. - Sources: official link

This week's trends

1. Fast, cheap models take center stage: Mercury 3 (diffusion architecture, 1,000+ tokens/s) and MiniMax M3.1 Flash (70% lower cost) launched on the same day (10-19), both pitching agentic and real-time use cases. 2. A wave of voice updates: Voxtral 3 (cloud plus voice agent SDK), the Polly Generative upgrade (cloud TTS) and QVAC TranscribePsy (on-device ASR) all shipped within one week, spanning cloud to phone. 3. Open weights keep advancing: Meta opened the 816B-parameter Muse Spark Open; Mistral released Timestral, a time-series foundation model (10-15, modified MIT license, https://mistral.ai/news/timestral); MiniMax has also announced it will open its weights. 4. MCP and the agent ecosystem: Sunra MCP brings multimedia models into existing assistants; Codex CLI keeps shipping frequent releases.