AI.info
AI Tools Updates — Page 83
What changed in the AI tools this week — dated, from the makers’ own changelogs.
- Opik — 2.2.96-6987
Hide and drop sampling parameters for OpenRouter.
- Opik — 2.2.95
Updated setup snippets, TS SDK OQL escape handling, Helm chart references, model prices file, and playground UI.
- Opik — 2.2.96-6983: [OPIK-8720] [BE] Serve a prompt's latest version from the index (#8825)
Serve a prompt’s latest version from the index.
- Opik — 2.2.96-6986
Keep Logs filters when returning to the page.
- Keploy — v3.6.107
Pairs each recorded response with its request and reports mocks left out.
- Keploy — v3.6.108
- llama.cpp — b11507
Added support for MoE cache across multiple GPUs.
- llama.cpp — b11512
DFlash now reads tied output weights from GGUF metadata and shares tied word embedding metadata.
- llama.cpp — b11511
Fixed CUDA MMQ out-of-bounds reads.
- llama.cpp — b11506: server : preserve context checkpoints across slot save/restore (#26004)
Server slot save and restore now preserve context checkpoints.
- llama.cpp — b11505
Fixed Vulkan TOP_K handling for +inf, NaN, and k = 1 on negative values.
- llama.cpp — b11510
Added a looped CUDA PAD kernel for more than 65,535 rows or slices.
- llama.cpp — b11509
Fixed the CCCL version guard for major version rollovers.
- llama.cpp — b11513: CUDA: improve top-k algorithm selection (#28713)
CUDA top-k implementation is selected by shape, with radix top-k for large row counts.
- JetBrains AI Assistant — Mellum2.1 Gets to Work: A Fast Open Model for Coding Agents
Released Mellum2.1, an open 12B mixture-of-experts model for coding agents under the Apache 2.0 license.
- Everlaw Ediscovery Platform — New Harvey + Everlaw Integration Brings Massive Scale to Legal AI
Everlaw and Harvey announced an integration connecting Harvey with Everlaw’s governed evidence layer for legal AI workflows.
- Everlaw Ediscovery Platform — Everlaw and Thomson Reuters CoCounsel Legal Launch New Partnership
Everlaw and Thomson Reuters announced a forthcoming CoCounsel Legal integration connecting matter evidence and legal research.
- Everlaw Ediscovery Platform — Everlaw Brings Frontier AI Closer to the Evidence with ChatGPT Plugin
Everlaw launched a ChatGPT Enterprise plugin connecting its governed evidence layer with ChatGPT Enterprise.
- Everlaw Ediscovery Platform — Everlaw and Microsoft Copilot Announce New MCP Integration (Available Now)
Everlaw's MCP integration with Microsoft Copilot is available, bringing Everlaw's evidentiary system of record to Microsoft 365.
- Krea AI — Workspace onboarding step for open-source usage reporting
Enterprise customers reporting open-source model usage now need to tell Krea which workspace they report from.
- n8n — n8n@2.42.5
Fixed the editor to show the Agent artifact while the builder creates it.
- n8n — stable
Fixed the editor to show the Agent artifact while the builder creates it.
- n8n — n8n@2.43.2
Fixed OpenAI-compatible Assistant model providers on self-hosted Assistant and showed the Agent artifact while the builder creates it.
- Ollama — v0.40.1
The server proxies cloud usage and balance APIs, and CLI onboarding no longer includes the account step.
- Hyperscience Platform — 43.2.3 (7 Oct 2026)
Machine Transcription uses configured Table Target Accuracy for table cells; reports treat missing output and blank Transcription QA as equivalent.
- Rox — Salesforce field mapping: You can now map Salesforce contact and lead fields into Rox, alongside existing account fields.
Map Salesforce contact and lead fields into Rox; the agent can use that context to personalize sequence emails.
- Gumloop — 12.1.0 · Tahsis
Tell Gumball what you want an agent to do and it builds or updates the agent, with its instructions, connectors, skills, knowledge, and triggers.
- OLMo — archive/speed-norms-interval
Gradient-norm bookkeeping now runs once per step.
- OLMo — archive/speed-full-window
The OLMo 3.5 recipes leave prefetch_keep_full off.
- OLMo — archive/speed-no-locks
OLMoOCR pages are rendered in helper processes using thread-local shard handles.
- OLMo — archive/speed-process-workers
Thread prefetch remains the default for mixed mid-training; process mode is available.
- OLMo — archive/speed-device-once: Stage the pixels to the device through a pinned buffer on a side stream
Image batches are staged through a reusable page-locked buffer and copied on a side stream.
- OLMo — archive/speed-compact-collator: Collate images compactly for the OLMoDDP multimodal recipes
An opt-in compact_images setting collates real image crops without padding to the batch maximum.
- Anima — Google Docs Markdown: Edit .md Files Natively
Google Docs began rolling out native Markdown editing, allowing users to open, edit, comment on, and share Markdown files.
- Hyperscience Platform — Removal of Python 3.12 support
Beginning in v44, Hyperscience no longer supports Python 3.12 in flows, including Code Blocks, Interpreter Blocks, and external Python packages.
- Coveo for Commerce — October 7, 2026 update
Managing source mappings is now a full-page experience, with dedicated tabs and an option to edit raw configuration with JSON.
- AssemblyAI — Fix: Clearer Error for Deleted API Keys
Requests made with a deleted API key now return an error that identifies the invalid key.
- Snowflake Cortex Analyst — Claude Haiku 5.5 model in the Cortex REST API (Preview)
Claude Haiku 5.5 model in the Cortex REST API.
- Tonic Textual — v494
Bug fixes and other internal updates.
- Rox — Salesforce field mapping
Map Salesforce contact and lead fields into Rox, and reference mapped fields in sequence instructions to personalize emails.