Search “Reasoning”

Fireworks Offers Free Training of Muse Glimmer 30B: Using SFT to Predict Kalshi Market Outcomes

Fireworks AI announced that Muse Glimmer 30B can be trained with SFT for free until November 30, and released a complete recipe for a prediction experiment using historical Kalshi market data. The fine-tuned Glimmer achieved 58.3% accuracy on unseen topics, 11.2 percentage points higher than Jev's 47.1%; in a historical backtest, a hypothetical $1,000 grew to $1,071 after accounting for simulated fees, while both base Glimmer and Jev lost money. Training used LoRA adapters and a K-way cross-entropy loss, ran through the training data only once, and separately used a calibration set to fit a temperature parameter for each question category.

MIT Course MAS.S62: Two Lectures on the Evolution of Harnesses and Multi-Agent Systems

Lectures 4-5 of the MIT course MAS.S62 "How to AI an AI" focus on Harness evolution, covering the evolution of harness components, memory, tools, and skills, as well as the evolution of planning, task decomposition, and multi-agent systems. The memory section uses a three-axis taxonomy to connect MemGPT, mem0, MEM1, and Reflexion/ExpeL, and uses CLM to unify context management as file operations; the tools section moves from Toolformer, ToolLLM, and CodeAct to Meta-Harness, which represents the entire harness as a file system and provides about 400 times the amount of feedback information of previous methods. The multi-agent section presents five empirical findings, including that MAS wins 81% on decomposable tasks and loses 70% on tasks with sequential dependencies, and that CORAL achieves new SOTA on 8 tasks, with multi-agent evolution improving over single-agent by 4%-20%.

English translation is being prepared.

English translation is being prepared.

How to Design AI Cost Control in Security Operations: Keeping Accuracy While Cutting Spending

A security operations team lead described how to control AI detection costs through tiered filtering; his team reduced AI detection costs in a Trust & Safety scenario to about $1 per day. The approach is to first use rules to filter out most obvious events, sending only a small number of events to the LLM, then have a lightweight model make an initial judgment and pass only low-confidence events to a stronger model. The team's tests found that for the target use case, the accuracy gap between the lightweight model and the frontier model was only 1% to 2%, while the frontier model's cost per token was about 5 times that of the former. The author also mentioned that his team wrote a prompt of more than 1,900 words for agent detection, and emphasized that cost should be taken into account at the design stage.

English translation is being prepared.

English translation is being prepared.

AWS Publishes Agent Registry Runtime Tool Discovery Practice

AWS introduces the Agent Registry runtime tool discovery approach: agents use hybrid semantic and keyword search to find approved MCP servers and A2A agents, then instantly build clients from descriptors to execute requests. The sample code is based on Amazon Bedrock AgentCore Gateway and runtime, integrates an order management MCP server and two A2A agents for pricing and customer service, and demonstrates a return and refund scenario using pure semantic matching. The registry returns only connection metadata; invocation credentials must be obtained separately through Cognito OAuth or IAM, and new capabilities require no changes to orchestrator code.

MiniMax-M3.1-Flash scores 42 on the Artificial Analysis Intelligence Index

MiniMax-M3.1-Flash scores 42 on the Artificial Analysis Intelligence Index, 13 points higher than MiniMax-M3, at less than half the latter's per-task cost. It scores 48% on Humanity's Last Exam and 22% on CritPt, both higher than GLM-5.3-Flash and Gemini 3.8 Flash, and slightly lower than DeepSeek V4.1 Flash 1017. The model has a 1M-token context window, supports text and image input, is priced at $0.15/$0.60 per 1M input/output tokens, and its weights are currently closed-source; MiniMax announced that it will open the weights on October 23.

How to run n8n AI workflows on NVIDIA DGX Spark without calling cloud APIs

n8n Blog published a tutorial on how to deploy a complete model stack locally on a single NVIDIA DGX Spark, enabling n8n workflows to perform transcription, OCR, document Q&A, and image generation without calling cloud APIs. The entire stack exposes an OpenAI-compatible endpoint through LiteLLM; the main model is the NVFP4 version of Qwen3.6-35B-A3B, with a single-user generation speed of 120 tokens/s and an aggregate of 359 tokens/s across eight users. The author's testing found that image generation reduces chat speed to about one-third, and recommends that small teams queue image tasks or schedule them outside working hours.

Surface Laptop Ultra review: NVIDIA RTX Spark's local AI performance

The author tested Microsoft's Surface Laptop Ultra equipped with the NVIDIA RTX Spark N1X SoC. The top configuration of this machine features a 20-core Grace Arm CPU, a 6144-core Blackwell GPU, and 128GB of unified memory, with a starting price of $2,599.99. In the UL Procyon AI text generation test, it scored 2703, about 33% higher than the IdeaPad Pro 5 with an Intel Core Ultra X9 388H; MLPerf Flux 2 Klein 4B image generation reached 26.97 images per minute, about 2.7 times that of the latter. The author also used a local Qwen3.8-27B to drive the Hermes agent, which completed extracting benchmark charts from Excel and generating a Word report in about five minutes.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Cerebras summarizes five orchestration patterns for multi-agent workflows

Cerebras published a blog post summarizing five patterns for building multi-agent workflows: parallel generation followed by human selection, swarm parallelism, staged parallelism, sequential handoff, and separation of building from verification. Using Codex with Figma MCP to replicate a website as an example, the article states that a single agent averaged 36.5 minutes per run, 12 human interventions, and a 100% failure rate, while the multi-agent setup took 5.2 minutes, required 2 interventions, and succeeded on the first attempt; in internal testing, the sequential loop reduced human interventions by 84.3%. The author argues that multi-agent setups expand the effective context window from about 200K to over 25M, and notes that Codex Spark's speed of about 1,200 tokens/second makes parallel and verification steps feasible.

Stanford's HomeBody connects GPT-6 Astra to the Unitree G1 to tidy up the kitchen

The Stanford team launched the HomeBody project, connecting GPT-6 Astra to the Unitree G1 to enter a kitchen and execute language instructions; it can gather coffee bags, discard designated cardboard boxes, and, when medication is not in view, find the drawer from memory and retrieve the medicine to hand to a person. The system first has Astra guide the G1 to explore an unfamiliar kitchen and save camera images, LiDAR, and joint poses as spatial memory, then reconstructs a digital kitchen in Isaac Sim, after which Astra calls reusable skills such as navigation, grasping, placing, and opening drawers, with the skill modules handling specific motion planning and feeding back results. The team says deploying to a new kitchen requires no additional training data collection or training of specialized motion policies, but perception, planning, and skill execution run on a single RTX 4090 laptop, Astra makes decisions remotely, the initial reconstruction involves preparation time and API costs, and overheating of the finger servos is also one of the current limitations.

Fireworks Evaluates DeepSeek V4 Pro Against Fable 5 and Kimi K3

Fireworks tested DeepSeek V4 Pro 0813 on four benchmarks: SWE-Bench Verified, LiveCodeBench v6, Aider Polyglot, and Terminal-Bench 2.1. It scored 95.2% on SWE-Bench Verified, higher than Kimi K3's 92.6% and Fable 5's 85.4%, and 92.0% on LiveCodeBench v6. On cost, DeepSeek V4 Pro spent $0.309 per task solved on SWE-bench, versus $0.808 for Fable 5; on LiveCodeBench the figures were $0.040 and $0.225, respectively. Its weakness is the multilingual breadth of Aider Polyglot, where it scored only 48.9% on Java, versus 74.5% for Fable 5. After oracle routing on a combined set of 409 tasks, accuracy was 92.4%, higher than the 86.6% achieved by Fable 5 alone, at a cost of $0.279 versus $4.510 per task, with 86.3% of tasks routed to V4 Pro.

Kimi K3 Complete Developer Guide: 2.8 Trillion Parameters, 1M Context, and Together AI Usage

Together AI has released the Kimi K3 developer guide, introducing Moonshot AI's 2.8-trillion-parameter open-weight model, which uses KDA hybrid linear attention, Attention Residuals, and Stable LatentMoE (896 experts with 16 activated), and supports a 1M token context and native visual input. The guide provides usage details for the OpenAI-compatible API, including reasoning_effort (low/high/max, default max), streaming reasoning_content, structured outputs, tool calling, and dynamic tool loading, and notes that sampling parameters are fixed at temperature 1.0 and top_p 0.95. Pricing is $0.30 per 1M tokens for cached input, $3.00 per 1M tokens for uncached input, and $15.00 per 1M tokens for output, with thinking tokens billed as output.