8 min read · 15 sources

MiniMax’s M3.1 Flash is the model to watch today: it cuts inference costs by 70%, runs nearly three times faster online, and pushes harder on coding, computer use and security. The company says its agent-focused model, derived from Space Bunny, processed more than 200 trillion tokens anonymously before launch.

That combination matters more than another leaderboard win. Agents burn tokens in loops, make tool calls and wait on sub-second responses, so latency and unit economics decide whether a workflow can run continuously or remains a demo. M3.1 Flash is available through MiniMax’s API, M Plan and MiniMax Code, with weights promised later.

Mercury 3 is trying to change the generation model itself. The diffusion-based LLM reportedly hit 1,068 tokens per second on Nvidia GPUs, alongside a 1-million-token context window. It also completed 50% of a 50-task Mind2Web web-navigation test, beating GPT-6.1-Sol under the same token-volume and agent-step constraints.

AWS’s push-based KV transfer cut time-to-first-token by up to 3× by moving cache blocks directly between accelerator memory.

MiniMax M3.1 Flash Targets the Agent Cost Curve

Source: minimax.io ↗

MiniMax M3.1 Flash is built for workloads where the model repeatedly plans, edits code, operates a computer and checks its own work. It scored 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld v2, 29.6 on ALE and 82% on CyberGym.

The more consequential claim is operational: MiniMax says online inference is nearly three times faster and usage costs are 70% lower. Those are vendor numbers, but they align with the direction agents need to move. The model weights have not shipped yet, so independent testing of the quality-cost trade-off still has to wait.

Mercury 3 Generates at 1,068 Tokens per Second

Source: inceptionlabs.ai ↗

Mercury 3 is a diffusion-based language model more than 40 times larger than Mercury 1. Inception Labs reports 1,068 tokens per second on Nvidia GPUs, with support for a 1-million-token context window.

Its price is $0.20 per million input tokens and $0.70 per million output tokens, temporarily cut by 50% to $0.10 and $0.35. Cached input tokens receive a 90% discount, which could make long, repeated agent contexts much cheaper.

The web-navigation result is the useful warning against treating raw speed as the whole story. Mercury 3 completed 50% of the Mind2Web test, outperforming GPT-6.1-Sol with equal token volume and agent steps. For latency-sensitive agents, throughput and task completion both matter; one does not guarantee the other.

Cloudflare’s Clairic Skips Token Generation Entirely

Source: blog.cloudflare.com ↗

Clairic is the first member of Cloudflare’s “System One” model class, available on Workers AI as @cf/cloudflare/clairic. Open weights are planned.

Instead of generating prose, Clairic extracts, calculates, estimates and forecasts structured values from unstructured input. Results are typed and carry confidence scores, while numerical outputs span an 80% range. The architecture can still be wrong, but it removes an entire class of malformed output and free-form hallucination.

For recurring jobs such as reading invoice totals, delivery dates, repair estimates and traffic forecasts, this looks like a cheaper alternative to a frontier-LLM call. The key operational question is whether its confidence and typed outputs map cleanly into existing validation and exception-handling paths.

Meta’s Prometheus Supercluster Is Learning to Be Boring

Source: engineering.fb.com ↗

Meta says the first phase of its Prometheus supercluster in New Albany, Ohio, is complete. It already contains hundreds of thousands of interconnected GPUs and is intended to grow toward roughly 1 million.

The cluster trained Meta’s Muse foundation models and Muse Wonder frontier models. Meta reports more than 85% goodput for pretraining and more than 90% for reinforcement learning, while mean time between interruptions improved from roughly 47,000 GPU-hours to more than 54,000 after fleet and firmware fixes.

That last metric is the real infrastructure story. At this scale, the useful unit is not peak FLOPS but the percentage of expensive accelerators kept on useful work. Closed-loop liquid cooling, fiber networking and fleet-level failure handling now matter as much as the chip procurement spreadsheet.

Benchmarks May Be Training the Industry to Optimize the Wrong Thing

Source: research.contrary.com ↗

The danger with benchmarks is not that models score badly. It’s that they score well, become training signals, and turn into objectives that no longer represent useful work. Contrary’s report on climbing the wrong hills argues that benchmark construction now directly shapes model behavior.

“Benchmaxxing” turns a test into a target. Teams can optimize prompts, tool use, data selection or post-processing until a score improves without making the underlying system more capable. Engineering leaders should keep benchmark gains separate from production evaluations, especially when vendors quote a single result without exposing the harness and token budget.

The IMF Got a First-Hand Warning About the AI Curve

Source: importai.substack.com ↗

Import AI 477 reproduces the Per Jacobsson Lecture delivered at the IMF meeting in Bangkok on October 17. Jacobsson argues that AI is already operating at the frontier of human science, sometimes autonomously, and that rapid capability growth is affecting economic and national-security planning.

The speech comes from someone who reported on data centers, worked at OpenAI and co-founded Anthropic. It is still a viewpoint rather than a measurement, but it provides a useful snapshot of how a frontier-model researcher is translating capability gains into a policy warning.

Exa 3.0 Makes Retrieval Cheaper and Contexts Smaller

Source: exa.ai ↗

Exa 3.0 claims first place across every benchmark it tested for coding, finance, deep research and company intelligence. Those are vendor-run results, but the release also exposes useful infrastructure numbers: Exa tracks more than 2 trillion URLs, serves 100 billion pages and indexes billions of documents.

Its new Dynamic Highlights feature uses 95% fewer tokens than sending full-page text, a meaningful reduction for agent pipelines that repeatedly fetch the same sources. Exa Lite costs $4 per 1,000 requests and shares Exa Auto’s index while applying less retrieval compute.

TorchNeuron Brings Native PyTorch to Trainium

Source: pytorch.org ↗

TorchNeuron is an open-source native PyTorch backend for AWS Trainium, entering public beta on October 20 under the Apache 2.0 licence. It supports eager execution, adaptive eager mode, torch.compile, the Neuron Kernel Interface, and Trainium2 and Trainium3.

The practical advantage is reduced platform friction. Teams can use familiar PyTorch model code, libraries and tooling instead of maintaining a separate accelerator workflow. Neuron Explorer also links torch.profiler traces to source lines and hardware, which should shorten the debugging loop when an AWS accelerator build behaves differently from CUDA.

Office Intelligence Keeps Evidence Attached to Long Jobs

Source: officeintelligence.github.io ↗

Office Intelligence is a research program for agents that work across environments, data and artifacts over long periods. Its example reconciliation workspace asks agents to prepare editable workpapers for deposits, card clearing and vendor credits while preserving source records, citing evidence and leaving unresolved differences visible.

The system declares 35 initial files and combines role guidance, operating procedures, project records and review templates. The useful pattern is continuity: the agent leaves artifacts and provenance for later review instead of producing a one-shot answer that disappears into chat history.

Anthropic and Nscale May Be Planning Another $45B Compute Deal

Source: techcrunch.com ↗

Anthropic and Nscale are reportedly in talks for another $45 billion compute agreement, after announcing a $45 billion deal a little over a month earlier. Both cautioned that discussions are ongoing and terms may change.

The scale points to an infrastructure strategy that extends far beyond buying capacity for one training run. It also raises financing questions as Anthropic prepares for an expected IPO this year and Nscale seeks a public listing despite investor pushback over its proposed $30 billion valuation.

HUMXN Is Buying Robot Training Data With Free Repairs

Source: tomshardware.com ↗

HUMXN is subsidizing plumbing, electrical, HVAC, air-duct and carpet services in Minneapolis, Chicago and Miami. Workers wear cameras that record repairs inside real homes, creating task data for training robots.

It is a practical way to obtain difficult embodied data, but it also raises consent and privacy questions that simulation does not: cameras observe private spaces, workers and AI labs have different incentives, and production models may learn from environments that were not represented fairly.

Meta’s Thin Orchestrator Deletes Work From the Router

Source: engineering.fb.com ↗

Meta’s 1,300-line thin orchestrator routes more than 20 data-pipeline workflows to 14 specialized agents without doing domain reasoning itself. Replacing one context-heavy agent with this design improved production reliability by isolating subsystem knowledge and reducing hallucinations.

A cited study found orchestrated systems reached 65.3% accuracy on 80-task batches, versus 16.6% for single agents, while using up to 65 times fewer tokens. The lesson is architectural: keep the router focused on sequencing and validation; keep domain decisions inside narrowly scoped agents.

Gemini Notebook Adds Review Before Rendering

Source: blog.google ↗

Gemini Notebook now lets users review and steer slide-deck and video plans before generation. Review controls expose slide-outline notes and video-producer notes, covering topic, visual style, ordering and format.

The rollout will reach all Gemini Notebook users globally over the following week, with an option to skip review for faster output. It is a small workflow improvement, but one that puts a human checkpoint before an expensive, visible artifact is rendered.

OpenAI Pushes Coding Agents Into Legacy Languages

Source: openai.com ↗

OpenAI’s report on low-resource programming languages says its models have improved at coding and software migration in COBOL, Fortran and Visual Basic. The release lacks a detailed benchmark breakdown, so the size of the improvement remains hard to judge.

The practical opportunity is huge in installed systems where modernization means translating or repairing old code rather than starting a new repository. Teams still need strong tests, because language familiarity and syntactically valid output do not guarantee preserved business semantics.

vLLM Pushes KV-Cache Blocks Directly Between Accelerators

Source: aws.amazon.com ↗

AWS added a push-based NIXL connector for vLLM’s disaggregated prefill/decode inference. It moves KV-cache blocks directly between GPU or Trainium device memory over RDMA-capable links such as EFA, avoiding host-memory staging and taking post-prefill coordination off the time-to-first-token path.

On EC2 P5en instances connected by EFA, the one-prefill/one-decode test produced up to 3× better TTFT and 30% better time per output token than pull mode. The gain was strongest at high request rates, exactly where coordination overhead tends to dominate disaggregated serving.