RL Briefing — October 20, 2026

October 20, 2026 · Daily briefing

The day's AI news, filtered through one question: what does this change for people who train agents? Five stories today — a cloud provider turns RL post-training into a managed service, a lab publishes the actual cost of a frontier RL run, a stealth-tested model cuts agentic task costs by 70%, open weights now ship with day-zero quantized builds, and agentic tools went on the offensive in a 100-site breach.

1. Together AI turns RL post-training into a managed service

Via RuntimeWire (announced on X by CEO Vipul Ved Prakash)

Together AI launched Together Post-Training, a managed service letting developers customize open models with supervised fine-tuning or full reinforcement learning. Developers define their data, rewards, environments, and training loop; Together runs the tuned GPU infrastructure, and customers keep the resulting weights or deploy them there.

Why it matters for RL: RL has always had a tooling gap — the loop (rollouts, reward, grading, training) is harder to run than it is to design. Post-training-as-a-service collapses that gap for teams without a training stack, which means more practitioners can actually iterate on reward design and environment quality instead of babysitting GPUs. Expect the differentiator to become who has the best environment and grader, not who owns the cluster.

2. Xiaomi's "You Only RL Once": the receipts on a $2.6M RL run

Via Towards AI (analyzing Xiaomi's MiMo-V2.6 technical report)

Xiaomi's new MiMo-V2.6 technical report is unusually concrete about what scaling RL actually costs: $2.6 million for the RL stage of the largest model, with training batches of roughly 25,000 agent trajectories and billions of tokens, plus a breakdown of where the compute went across rollout, grading, and training. The report names its own experiment section "You Only RL Once" — a nod to how high-stakes the bet is.

Why it matters for RL: as pretraining returns flatten, the debate is whether RL can keep delivering gains — and this report says yes, but with a price list: scaling RL is three problems at once (more compute, richer environments, smarter grading). The public cost breakdown is a gift for anyone planning an RL run: grading and rollout infrastructure, not just the optimizer, dominate the bill. Budget your verifier before your model.

3. MiniMax's M3.1 Flash cuts agentic task costs 70%

Via Technology.org

MiniMax released M3.1 Flash, ending the mystery of "Space Bunny" — an anonymous model that climbed developer leaderboards while processing over 200 trillion tokens in stealth. The 428-billion-parameter model prices at $0.15 per million input tokens and claims a 70% cost cut on finished agentic tasks versus its predecessor, with output above 200 tokens per second. Artificial Analysis scores it 42 on its Intelligence Index, 13 points above M3.

Why it matters for RL: agent economics are task economics, not token economics — MiniMax's own Artificial Analysis comparison shows the per-task cost drop (57% there) matters more than the per-token number. Stealth leaderboard testing is also becoming the standard model-release playbook, which means eval environments keep inheriting "solved by something anonymous" baselines. Build your eval budget around finished tasks, not token burn.

4. Open weights now ship with day-zero GGUF quants

Via AI News Brief

Meta's 816-billion-parameter Muse Spark Open shipped October 15 with Unsloth's GGUF quants already live — made possible, per Unsloth, by day-zero access from Meta. Cloudflare did the same with bartowski on its Clef model October 7. Labs are now coordinating open-weight releases with the quantization community before launch, not days after.

Why it matters for RL: open weights are the raw material for RL practitioners — the base you fine-tune, distill from, or roll out against in your training harness. Day-zero quants mean the community can experiment on consumer hardware within hours of release, which accelerates the whole downstream loop: distillation targets, reward-model bases, and rollout models all get better faster. Watch for which labs make this coordination standard practice.

5. Agents go on the offensive: agentic tools used in a 100-site breach

Via The Neuron (Reuters reporting)

Reuters reported that a hacker used agentic AI tools to breach more than 100 company sites — including John Deere and SharkNinja — with little human intervention. Meanwhile, Microsoft rolled out enterprise controls that assess agent risk and can restrict or shut down risky agents, and OpenCode disclosed that stolen Console API keys are already being used to run up real charges on developers' accounts.

Why it matters for RL: this is the mirror image of training agents to do useful work: the same tool-use and multi-step planning we optimize for is now carrying out attacks at scale. Two lessons for the lab — treat eval sandboxes and stored credentials (API keys especially) as the attack surface they now demonstrably are, and expect "resistance to misuse" to become an explicit eval axis. Microsoft's monitoring tooling hints at where the field is heading: capability gated by continuous risk assessment, not just a one-time safety check.

The one-line version

RL as a service: Together Post-Training lets anyone run custom SFT/RL loops on open models — the moat moves to environments and graders.
Scaling has a price tag: Xiaomi's MiMo-V2.6 puts a $2.6M, 25,000-trajectory figure on a frontier RL run — budget the verifier first.
Agent costs keep falling: MiniMax's M3.1 Flash cuts finished agentic-task costs 70% — task economics beat token economics.
Open weights arrive ready to run: Meta's 816B Muse Spark Open ships with day-zero GGUF quants — distillation and RL targets in hours.
Agents on offense: agentic tools breached 100+ sites with little human intervention — sandboxes and stored keys are the attack surface.

The RL Briefing publishes daily on rlenvlab — the day's AI news through the lens of people who train agents. More from the lab notebook →