Parallel for Finance is announced as the most accurate search API on FinSearchComp's hardest tier (T3) with up to 2× lower cost than Exa and Perplexity. It also reports cost and accuracy results on FrontierFinance and a private PWS-Finance benchmark.
Key points
Accuracy: Parallel is the most accurate search API on FinSearchComp T3, up to 2x lower cost than Exa and Perplexity.
Benchmark: On the private PWS-Finance benchmark, Parallel leads, beating OpenAI's native search at 30% lower cost.
Sources
zml/llmd 20261029.0 release with Mistral Large 4, GLM 5.2/5.3, LFM 3
zml/llmd 20261029.0 adds Day 0 support for Mistral Large 4, GLM 5.2 and 5.3 series, and LFM 3 in BF16, FP8 and NVFP4. The release includes docker run commands for each model and platform.
Key points
Day 0 support for Mistral Large 4, a 1-trillion-parameter multimodal model with three reasoning modes and native speculative decoding via --speculative-drafter-model=native.
Docker run examples for LFM 3 on NVIDIA, AMD, Intel, Google TPU, and Apple Metal, each with specific flags and model URLs.
Sources
- Original post
- zml/llmd 20261029.0 - Linked page
Serval raises $150M Series C at $3B valuation, expands team 10x
Serval announced a $150M Series C funding round at a $3B valuation, bringing total funding to $277M. The company reports a 10x team growth in the last year and lists several enterprise customers.
Key points
Funding: $150M Series C led by Sequoia, valuation $3B, total funding $277M.
Growth: Team size increased 10x in the past 12 months; customers include Fox Corporation, WPP, Live Nation, Lucid Motors, and Ramp.
Sources
Claude self-assessment flags misuse after response
Claude self-assessment runs right after Claude responds, giving a preliminary safety assessment. It has been used since June and has surfaced misuse that other safeguards missed.
Key points
Self-assessment runs right after Claude responds, giving a preliminary safety assessment of the conversation.
Used since June, it has surfaced misuse that other safeguards weren’t designed to catch.
Sources
joey-1 2B released, fastest decision model
joey-1 2B is a new open-source decision model described as the fastest of its size. It is offered through Model APIs, a dedicated deployment, and a fine-tuning cookbook.
Key points
Release: joey-1 (2B) is the third open-source decision model from the team.
Availability: joey-1 can be accessed via Model APIs, dedicated deployment, and a cookbook for fine-tuning on Training Jobs.
Sources
- Original post
- baseten.co - Linked in the post
Scira now uses YouDotCom Web Search API with multi-passage citations
Scira now runs on the YouDotCom Web Search API by default and cites several extracted passages per page. It also returns fresher news, date-bounded Lookouts, and images in the same response.
Key points
API: now runs on the YouDotCom Web Search API by default
Citations: every web-mode answer cites from several extracted passages per page
Content: includes fresher news, date-bounded Lookouts, and images in the same response
Sources
- Original post
- scira.ai - Linked in the post
Phonic-1 2nd on MIVAS Bench, first non-big-lab top-3
Phonic-1 placed 2nd on the MIVAS Bench, becoming the first model not from OpenAI, Google DeepMind, or SpaceX Super Intelligence in the top 3. The team highlighted that building strong S2S models is not limited to top model labs.
Key points
2nd place on MIVAS Bench, beating Grok 3.0
First model not released by OpenAI, Google DeepMind, or SpaceX Super Intelligence to reach MIVAS top 3
Sources
Scira default search provider changes to YouDotCom
Scira now uses @youdotcom as default search provider. Every web-mode answer across 5M+ searches now cites several extracted passages per page and includes fresher news, date-bounded Lookouts, and images in the same response.
Key points
Default search provider: @youdotcom Web Search API.
Web-mode answers cite multiple passages per page and add news, date-bounded Lookouts, and images.