AI News: April 25, 2026

1. DeepSeek Previews V4-Pro and V4-Flash, Closing Gap to Frontier Models

DeepSeek. DeepSeek released open-weights previews of V4-Pro (1.6T params total / 49B active) and V4-Flash (284B / 13B active), both 1M-context MoE models, on April 24. V4-Pro is now the largest open-weights model and falls only “marginally short” of GPT-5.4 and Gemini 3.1-Pro on reasoning — a gap DeepSeek frames as 3 to 6 months. Pricing of $0.14/$0.28 per million tokens for Flash and $1.74/$3.48 for Pro makes Flash the cheapest small model and Pro the cheapest larger frontier model, with optimization targeting Huawei chips. Source

2. Cognition in Talks to More Than Double Valuation to $25B

Cognition. Bloomberg reported that Cognition is in early talks to raise hundreds of millions at a $25 billion valuation — more than double the September 2025 round at $10.2 billion. Devin grew ARR from $1 million in September 2024 to $73 million by June 2025, and the round comes shortly after Cognition’s Windsurf acquisition. The ask reflects investor demand for AI-coding leaders amid pressure from Anthropic, OpenAI, and Cursor. Source

3. ComfyUI Raises $30M at $500M Valuation as Creators Push for Granular Control

ComfyUI. ComfyUI announced a $30 million round at a $500 million valuation led by Craft, with Pace, Chemistry, and TruArrow participating, bringing total funding to $48 million. The node-based open-source workflow tool now claims more than 4 million users, 60,000+ community nodes, and 150,000+ daily downloads. It also powered SVEDKA’s 2026 Super Bowl ad — the first largely AI-generated Super Bowl spot — pointing to the growing professional adoption of node-based generative pipelines. Source

4. Omni Raises $120M Series C at $1.5B for Enterprise AI Analytics

Omni. The ex-Looker co-founders’ BI-and-AI-analytics startup raised a $120 million Series C at a $1.5 billion valuation, led by ICONIQ with Theory, First Round, Redpoint, and GV. The round, which includes a $30 million employee tender, represents a 2.3x step-up from a $650 million valuation in March 2025 and is tied to 4x year-over-year revenue growth driven by BI consolidation, AI adoption, and enterprises building on Snowflake, Databricks, and BigQuery. The round positions Omni as the leading challenger to Looker and Mode in semantic-layer-driven analytics. Source

5. Cloudsmith Raises $72M Series C to Govern AI-Generated Software Supply Chains

Cloudsmith. The Belfast-based artifact-management startup raised $72 million in Series C funding led by TCV with Insight Partners and existing investors, taking total funding past $100 million and pushing valuation toward unicorn status. The pitch: as AI-generated code floods enterprise repos, Fortune 500 buyers need package-level governance — provenance, licensing, vulnerability gating — that legacy artifact tools can’t supply. The round signals continued investor appetite for the security-and-supply-chain layer beneath AI coding agents. Source

6. Orkes Lands $60M Series B for Agentic and Durable Workflow Orchestration

Orkes. The Conductor-derived workflow orchestration platform raised $60 million in Series B funding led by AVP, with Prosperity7, Nexus, Battery, and Vertex US participating, bringing total funding to roughly $90 million. Customers include Twilio, LinkedIn, Quest Diagnostics, United Wholesale Mortgage, Naveo Commerce, and Woodside Energy, as enterprises increasingly fold human-in-the-loop, agent, and microservice steps into a single durable workflow runtime. The round arrives alongside renewed interest in Temporal-style durable execution as a substrate for production agent systems. Source

7. Aaru Raises $80M Series A for Synthetic Consumer and Political Research

Aaru. Aaru, which uses agent simulations to model populations and predict consumer and political behavior, raised an $80 million Series A on April 23, bringing total reported equity funding to $88 million. The round adds capital to a category trying to compress traditional polling and market-research cycles by simulating responses instead of surveying humans. Whether simulated panels can substitute for live respondents on questions of taste and intent remains the open empirical question — the round suggests at least some buyers think they can. Source

8. PrismML’s Bonsai 8B Is a 1-Bit LLM That Runs Natively on iPhone 17 Pro Max

PrismML. AlphaSignal’s April 24 piece spotlights Bonsai 8B, a natively 1-bit LLM (every weight ±1) that fits in 1.15 GB and runs at roughly 44 tokens/sec on iPhone 17 Pro Max, 131 tokens/sec on M4 Pro Mac, and 368 tokens/sec on RTX 4090. It is roughly 14x smaller than 16-bit 8B peers, claims 4-5x better energy efficiency, and remains competitive on benchmark averages. The release is a meaningful data point for on-device AI: 1-bit native quantization-aware training appears to have closed enough of the quality gap that consumer hardware can now host capable assistants entirely offline. Source