DeepSeek V4 Is Coming: Pro and Flash Versions Emerge, Still Per-Token

Internal sources point to imminent release of DeepSeek V4 lineup with distinct Pro and Flash tiers, maintaining per-token pricing despite industry shifts toward subscription models.

The next generation of DeepSeek models appears close to launch, with model identifiers deepseek-v4-pro-202606 and deepseek-v4-flash-202605 surfacing in community discussions. If confirmed, the Pro variant would represent DeepSeek's most capable reasoning model to date, while the Flash version targets cost-efficient inference at scale. The per-token pricing strategy, however, runs counter to a growing trend among frontier labs toward flat-rate or subscription access. DeepSeek's decision to stick with metered billing likely reflects confidence in its inference cost structure, which benefits from aggressive hardware optimizations and a vertically integrated stack. Observers note that V3's release reshaped the open-weight frontier landscape, and V4 could trigger another wave of price competition across the API inference market, particularly if the Flash variant delivers competitive quality at a fraction of current frontier pricing.

Sakana AI Brings 11 Papers to ICML 2026, From Multi-Agent to Black-Box Optimization

Sakana AI will present a substantial body of work at ICML 2026 in Seoul, spanning multi-agent coordination, evolutionary algorithms, and foundation model techniques. Among the highlights, the paper "Bridging Spherical Black-Box Optimizers" proposes a unifying framework for connecting disparate optimization algorithms on spherical manifolds, aiming to close the gap between theory and practical deployment in high-dimensional model tuning.

Original Scaling Laws Paper Had a Bug, Leading to Years of Compute Waste

A researcher has revealed that a bug in the foundational scaling laws paper produced systematically wrong conclusions, causing labs to train models far larger than necessary for their data budgets. The bug meant optimal model sizes were underestimated, leading companies to burn enormous compute on oversized, undertrained models. This revelation comes before the industry even started properly accounting for inference cost in the total cost equation, compounding the inefficiency.

Claude Fable 5 Reasoning Distilled into Qwen3-4B with 100% Self-Consistency

A team from the University of Waterloo distilled 2.3 million reasoning traces from Claude Fable 5 into the compact Qwen3-4B model, achieving perfect self-consistency across 512 samples with zero-bit output errors. The result demonstrates that frontier-level chain-of-thought reasoning can be effectively transferred to much smaller, open-weight models through careful distillation, dramatically reducing inference costs while preserving output quality.

LongCat Weights Released: Largest Model Pretrained on Non-Western Chips

The LongCat model weights are now available on Hugging Face, representing the largest known pretraining run executed on non-Western hardware. The release offers a rare window into the capabilities of Huawei's chip stack at scale, decoupled from Meituan's data expertise. Evaluators are expected to scrutinize not just the model's benchmark scores but what the results say about the maturity of the underlying silicon ecosystem. The open-weight release invites the global research community to independently assess both the model quality and the hardware stack that produced it.

Compute Is Easier to Acquire Than Frontier Models, Two Years of Evidence Shows

A striking observation after two years of AI industry evolution: the path from having no compute to holding a solid chunk of global capacity is far shorter than the path from having abundant compute but no frontier model to having both. The implication is clear: algorithmic and architectural breakthroughs remain the true bottleneck. Hardware procurement, while capital-intensive, can be solved with money and partnerships. Building a genuinely competitive foundation model requires something far harder to buy: research talent, training infrastructure expertise, and the institutional knowledge that only comes from repeated, expensive failures.