Reflection AI released Beam, a 501B1-parameter MoE model, with open weights and a score of 80.1 on Terminal-Bench v2.1. It lands as open-weight models carry most of the token traffic.
Vercel's gateway data puts open-weight output at roughly 8.4 times less than closed models. Open models' share of revenue there is climbing about six points every month, so the cost gap is already shifting real spending.
SuperLocalMemory grades each recall and keeps replaced facts on an undoable record. That aims at the costliest multi-agent coding failure, an agent acting confidently on outdated context, and the free local core makes a trial cheap.
UCLA's PhoneBot is a 19-inch bipedal robot whose torso is an Android phone. The phone's camera, motion sensors and processor supply the vision and balance, which keeps the hardware cheap.
Reflection AI posted Beam's weights under Apache 2.0; it scores 80.1 on Terminal-Bench v2.1, below GLM-5.2 (81.0) and Kimi K3 (88.3).
Why it mattersBecause Beam's weights are under Apache 2.0, teams can self-host and fine-tune it commercially, though agentic coding work that demands the best Terminal-Bench v2.1 results still favors Kimi K3.
Continue readingThe Nasdaq fell 2.15% and the S&P 500 lost 1.09% while WTI crude rose 3.51%, leaving oil as the only listed asset to gain on the session; Bitcoin was flat, Ether slipped 0.58%, and the crypto Fear & Greed index read 53 (Neutral).
A weekly operator deep dive, the full Dojo Pro prompt packs, and the complete prompt database. Founding members lock the launch price forever.
Join the founding list →A featured guide covers Google's newly opened Gemini 4 Argon, a model built for long, multi-step work that current setups struggle to finish. It argues that most daily tasks do not need the stronger model and lays out a decision process for choosing between Argon, Flash and Flash-Lite.
Muse Spark Open reaches Macs as Meta released weights optimized for Apple silicon, but running it requires at least two top-spec M5 Ultra Mac Studios. Meta AI chief Alexandr Wang told Axios the two companies are 'playing nice again.'
Open Models: 71.5%2 of Tokens, 23%3 of Spend
Read full analysis
Tool of the day
7.413/ 10
Every recall from SuperLocalMemory comes back graded, so a coding agent about to bluff about your staging port gets caught before it acts.
SuperLocalMemory grades each recall and keeps replaced facts on an undoable record, which targets the costliest failure in multi-agent coding: an agent acting confidently on stale context. The free local core means a trial costs one install, and developers running two or more coding agents on one repo should take it this week. Keep it a personal tool until the paid plans carry real numbers.
CtxWise tree-shakes MCP tool schemas so your coding agent loads only the definitions it needs, which lowers context usage. It also adds firewall controls that restrict which tools an agent can reach. Run one typical task with and without it, compare the token counts, then use the firewall to block any MCP servers that task never needs.
Try itPaveo sits between you and agents such as Claude Code, Codex, and Cursor, inspecting the shell commands and file writes they issue. Every action goes into a hashed audit log, so later edits to the record are detectable. Turn it on for agents that touch production config or deployment scripts, and review the log after each session to see what actually ran.
Try itAuxx.ai is an open-source support platform. Its Kopilot assistant uses text classifiers to sort incoming tickets and drafts context-aware replies for agents to approve. Because the code is open, small support teams can self-host it and inspect how tickets are routed before letting it draft responses to real customers.
Try itEverygen AI, previously called Viewmax, creates both video and image content for marketing, social, and other uses. Keep your still images and short clips in one place by drafting a campaign's visuals and companion video together. Then compare the output against your current workflow to see whether it saves a hand-off between separate tools.
Try itLarge language models have evolved from single-turn chatbots into interactive agents capable of long-horizon reasoning.
- • EnvScaler [Song et al., 2026] constructs executable environments with verifiable outcomes and is the strongest environment-scaling baseline in our tables; because the ORPO stage of Agent-Reflex reuses its RL task pool, it is the reference for attributing...Read the paper
The paper comes from the Firecrawl Research Index: arXiv, indexed in the last 7 days and submitted this month or last, picked for today's lead. Quotes are verbatim.
Run Beam on your own terminal tasks. Pull the Apache 2.0 weights and replay ten real shell-heavy jobs from your backlog. Log pass rates next to your current model, since a public leaderboard score will not match your repo.
Price a routing split. Pull a week of gateway logs and tag each call by difficulty. Send the routine, high-volume traffic to an open-weight model and keep a closed model for the hard cases, then compare the bill.
Trial SuperLocalMemory on one multi-agent repo. Start with the free local core and review how it grades recalls and logs replaced facts. Roll back one deliberately stale fact to confirm the undo record works.
Beam arrives with a permissive license just as gateway data shows open-weight models taking a growing share of paid usage, which makes a self-hosted trial a practical question rather than a hobby. The benchmark order still matters for hard agentic coding, so judge Beam on your own workloads and not on the headline score. Pair that test with a memory layer such as SuperLocalMemory so stale context does not skew the results. Start with a small routed slice of traffic this week and expand only where the numbers hold.
Method names the computation path behind the row, from a closed vocabulary: as published, yfinance close, 03B search corroboration, 04R dossier scoring, not recorded.
A figure is money, a percentage, a spelled magnitude, a multiple, a rate carrying its unit, a score over its denominator, a thousands separated number or an index close. Bare integers and years are not counted, so a phrase like "over 100 integrations" carries no mark. A chart's own scale line is the axis of the level above it and is counted once, with that level.
205 editions so far, one a day. Unsubscribe anytime.
Forward this to one operator you work with. Your referral link is in every email; milestones at koda.community/refer.