ModelsSolar Pro 3: Specs, Retirement Date and Moving to Solar Pro 4
Upstage retires Solar Pro 3 on 30 October 2026. What the 102B MoE offered, what changes when you move to Solar Pro 4, and how to migrate without surprises.
ReadNotes on models, agents, and what inference costs
Practical writing for developers building with large language models — how they work, how to pick one, and how to keep the bill predictable.
Picking between models — and testing them on your own work.
ModelsModelsSolar Pro 4: Upstage's Agent Model for Korean, English, Japanese
Solar Pro 4 is Upstage's flagship for agentic work, with a 512K window, 128K output and a three-language focus. What is disclosed, what is not, and how to test it.
Read
ModelsSonar Deep Research: Budgeting a Report, Not a Reply
Sonar Deep Research runs many searches and bills on five meters at once. How to estimate the cost of a report, handle multi-minute calls, and when to use something else.
Read
ModelsSonar Pro Guide: What the Higher Price Actually Buys
Sonar Pro lists at three times Sonar's input price and fifteen times its output price, yet real bills differ far less. Where the money goes and when Pro is worth it.
Read
ModelsSonar Pro Search: Multi-Step Search Priced per Call
Sonar Pro Search is Perplexity's agentic search mode, sold as its own model id on OpenRouter. How its per-request fee works and how to tell whether multi-step search pays.
Read
ModelsSonar Reasoning Pro: Search Plus Reasoning, Mid-Migration
Perplexity's search-grounded reasoning model, how its two-part bill works, and what the end of Sonar Chat Completions support in September 2026 means for it.
ReadModelsSpace Bunny Alpha: How to Evaluate an Anonymous Model
Space Bunny Alpha was a free stealth model with a 1M window, since revealed as MiniMax M3.1 Flash. What the listing said, and how to test a model like it.
Read
ModelsStep 3.5 Flash: 196B Open Weights That Fit in 128 GB
StepFun's Step 3.5 Flash activates 11B of 196B parameters, ships under Apache 2.0 and runs quantised on one 128 GB machine. What the design buys and what it costs.
ReadModelsStep 3.7 Flash: Vision, Effort Levels and a Higher Output Price
Step 3.7 Flash adds a vision encoder and three reasoning levels to StepFun's 11B-active MoE. What changed from 3.5 Flash, what it costs, and when the upgrade pays.
ReadModelsTernary Bonsai 2 27B: A 27B Model in Under 6 GB
PrismML squeezed Qwen3.8-27B to 1.72 bits per weight and kept most of its quality. How ternary weights work, what they cost you, and what it takes to run.
ReadModelsTrinity Large Thinking: Arcee's 400B Open Reasoning Model
Arcee's Trinity Large Thinking is a ~398B MoE with 13B active, built for agent loops. Its licence changed after launch, and its reasoning tokens need careful handling.
ReadModelsUI-TARS-1.5 7B Guide: A Small Open Model That Drives a GUI
ByteDance open-sourced a 7B screen-operating agent under Apache-2.0. What it scores, how its click coordinates work, and how to run it without regretting it.
ReadShowing 313–324 of 446 articles