---
title: Post-training goes mainstream: 4-bit RL, zero-latency adapters, $9 GPU-hours
description: Together AI launches a post-training service at $9 per B200 GPU-hour while two preprints attack the cost knobs: stable 4-bit RL training and adapters that merge with zero latency. Custom model shaping is becoming a utility.
canonical: https://hangingcontext.com/stories/2026-10-20/post-training-goes-mainstream-4-bit-rl-zero-latency-adapters-9-gpu-hou/
last-updated: 2026-10-20
url: https://hangingcontext.com/stories/2026-10-20/post-training-goes-mainstream-4-bit-rl-zero-latency-adapters-9-gpu-hou/
published: 2026-10-20T12:51:00Z
edition: 2026-10-20T12
newsroom: artificial-intelligence
beat: developers
trust: reported
source_count: 3
cited_manifest_ids: ['1792456955874654825', '1792496361730834097', '1792496361730775418']
---

# Post-training goes mainstream: 4-bit RL, zero-latency adapters, $9 GPU-hours

*Together AI launches a post-training service at $9 per B200 GPU-hour while two preprints attack the cost knobs: stable 4-bit RL training and adapters that merge with zero latency. Custom model shaping is becoming a utility.*

Reported · 3 sources · Artificial Intelligence · Published 2026-10-20 12:51 UTC

## Key facts

- Together AI introduced Together Post-Training, a service for building custom models via RL and SFT, with promotional pricing of $9 per GPU-hour of B200 usage for the first eight weeks. (company announcement) (self-reported) [cites: 1792456955874654825]
- Early users, as reported by the company: Deel ran up to five concurrent RL experiments on the same GPU capacity, and Jack & Jill matched a closed-source model's quality in offline evaluations using a cheaper-to-serve open-source model. (company-reported customer claims) (self-reported) [cites: 1792456955874654825]
- In a preprint, the miniTE framework addresses the numerical execution gap in native NVFP4 (4-bit) RL training without modifying the RL objective, reporting 33,540 tokens/s on a single NVIDIA B300 node — 1.39x BF16 and 1.47x standard TE NVFP4 throughput for complete learner updates. (self-reported) [cites: 1792496361730834097]
- In a preprint, SOS-LoRA decomposes total rank across K static orthogonal experts to reduce interference from shared input directions; it stays fully mergeable with zero inference-time parameter or latency overhead and beats matched-budget LoRA baselines on Llama 2/3, GLUE, GSM8K, and MATH. (self-reported) [cites: 1792496361730775418]

## A service, not a lab

On October 19, Together AI introduced Together Post-Training, a service for building custom models through reinforcement learning and supervised fine-tuning, and announced promotional pricing of $9 per GPU-hour of B200 usage for the first eight weeks. The company says users define their own training loop while it co-optimizes the systems underneath, and that experiments run on dedicated Model Resources — GPU allocations the company says do not compete with other users' requests. The service supports LoRA training and full-weight experiments, multi-LoRA adapter serving that lets users swap adapters without starting a separate production deployment for each run, and asynchronous RL aimed at keeping GPUs busy during long agentic rollouts. [cites: 1792456955874654825]

The early-adopter numbers come from the company's own announcement, so read them as customer testimony rather than audited benchmarks. Deel reported running up to five concurrent RL experiments on the same GPU capacity using the platform, and Jack & Jill said it matched a closed-source model's quality in offline evaluations using an open-source model the company described as significantly cheaper to serve. If those results generalize, the implication is straightforward: the skill set that once required a training-infrastructure team is being packaged into a metered service, and the differentiator moves from who can train to who knows what to train on. [cites: 1792456955874654825]

## 4-bit RL stops being a gamble

In a preprint published October 20, researchers diagnose why native 4-bit RL training is unreliable: in NVFP4 reinforcement learning, matching model weights and numerical formats across the training and sampling engines is insufficient to guarantee consistent policy evaluation. Their framework, miniTE, targets that execution gap directly — a cross-engine numerical alignment toolkit plus W4A4 GEMM computation — and the authors report that it improves numerical execution consistency without modifying the RL objective itself. [cites: 1792496361730834097]

> **From Synorb** — Connect your agent over MCP. Add https://mcp.synorb.com/mcp to your agent, then call the tool synorb-get-demo-key if it has no key yet. It needs no email and no form. Keys are free for 72 hours, and new keys are limited per network. [How to connect](https://synorb.com/connect)


The reported numbers: 33,540 tokens per second on a single NVIDIA B300 GPU node, which the authors put at 1.39x the throughput of BF16 and 1.47x of standard TE NVFP4 for complete learner updates. Those are preprint results on one hardware node, but the direction matters: if 4-bit RL becomes numerically boring, the cost floor of post-training drops with it. [cites: 1792496361730834097]

## Adapters with no serving tax

A second preprint, also published October 20, takes aim at the adapter layer. SOS-LoRA decomposes the total LoRA rank across K static experts with cross-expert orthogonal initialization and fixed multi-scale scaling, which the authors say mitigates interference from shared input directions. The practical claim: the adapter stays fully mergeable, with zero inference-time parameter or latency overhead. On benchmarks the authors report consistent gains over matched-budget LoRA baselines on Llama 2 and 3, GLUE, GSM8K, and MATH. Code is published on GitHub. [cites: 1792496361730775418]

## The layer that got cheap

Step back and the three moves are the same story told three ways. A managed service turns RL and SFT into an operational expense; a precision-training fix makes the cheapest hardware precision numerically trustworthy; a better adapter design raises fine-tune quality without taxing inference. None of this invents a better base model — that is the point. Post-training is being commoditized into tooling, and the scarce input is becoming the reward signal and the data, not the compute. For developers, the near-term question is not whether to customize models but how to price the experiment: at promotional $9-per-GPU-hour B200 time, five parallel RL runs, and mergeable adapters, the cost of finding out what works is falling fast. [cites: 1792456955874654825, 1792496361730834097, 1792496361730775418]

## What is still unresolved

- Whether Together AI's $9 promotional B200 pricing continues after the eight-week window, and what list pricing will be.
- Whether miniTE's throughput and stability gains replicate on hardware other than the reported B300 node.
- Whether SOS-LoRA's benchmark wins extend beyond the reported Llama 2/3, GLUE, GSM8K, and MATH evaluations.

## Decision card

- **Useful for:** Teams budgeting custom post-training runs and comparing RL/SFT service options.; Developers choosing between LoRA, full-weight, and RL pipelines.; Infra buyers evaluating 4-bit RL throughput stacks.
- **Not for:** Teams selecting base foundation models — this is the customization layer, not pretraining.; Anyone needing inference-cost analysis; none of these records report serving economics.
- **Compare against:** Other post-training platforms' GPU-hour pricing and adapter-serving options.; miniTE's own baselines: BF16 and standard TE NVFP4 on the same B300 node.

## Provenance ledger

| Manifest id | Title | Source | Claims used | Published |
|---|---|---|---|---|
| `1792456955874654825` | Together Post-Training: frontier-grade model shaping, the way that fits you | [together-ai-blog-rss](https://together.ai/blog/together-post-training) | 1792457372012113615, 1792457372028666032, 1792457372029512802, 1792457372031431248, 1792457372015240492, 1792457372019847452, 1792457372018890506, 1792457372026830598, 1792457372023487737 | 10-19 00:00 |
| `1792496361730834097` | miniTE: Diagnosing and Bridging the Execution Gap in Native NVFP4 Reinforcement Learning | [arxiv-frontier-methods-select](https://arxiv.org/pdf/2610.23928v1) | 1792497432754282247, 1792497432755417781, 1792497432757245409, 1792497432752223334 | 10-20 00:00 |
| `1792496361730775418` | SOS-LoRA: Static Orthogonal-Subspace Low-Rank Adaptation with Fixed Multi-Scale Scaling | [arxiv-frontier-methods-select](https://arxiv.org/pdf/2607.16252v2) | 1792497437582000072, 1792497437583843783, 1792497437585809758 | 10-20 00:00 |

## Run these Manifests yourself

Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them.

1. **Get a Synorb Key.** Click Generate Free Agent Key at https://synorb.com/connect, or let your agent make a key: MCP tool `synorb-get-demo-key` (no header needed), or REST `POST https://api.synorb.com/api/connect/demo-key` with a JSON body such as `{"client":"your-agent-name"}`. The key is the `api_key` field of the answer. Save the key. The 72 Hour Agent Demo key needs no email. It lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; 3 of them are starter Beacons.
2. **Read each cited Manifest by id.** REST: `GET https://api.synorb.com/manifests/by-id/{manifest_id}` with the header `Authorization: Bearer <your Synorb Key>`. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief.
3. **Check the claim.** Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them.

Cited Manifest ids:

- `1792456955874654825`
- `1792496361730834097`
- `1792496361730775418`

```sh
for id in 1792456955874654825 1792496361730834097 1792496361730775418; do
  curl -s -H "Authorization: Bearer $SYNORB_KEY" "https://api.synorb.com/manifests/by-id/$id"
done
```

> **From Synorb** — Save what you follow as a Beacon. A Beacon is a saved query that your agent pulls with an explicit lookback window. Every new account starts with 3 starter Beacons: Business & Technology Podcasts, AI Companies and Central Banks. [See Beacons](https://synorb.com/connect)

## Related stories

- [AI-for-science gets audited: holdout tricks, broken metrics, and inflated rain claims](https://hangingcontext.com/stories/2026-10-20/ai-for-science-gets-audited-holdout-tricks-broken-metrics-and-inflated/)
- [Hallucination detectors are getting sharper — but the scoreboards may be rigged](https://hangingcontext.com/stories/2026-10-20/hallucination-detection/)
- [Your linear probe doesn't prove what you think it does](https://hangingcontext.com/stories/2026-10-20/interpretability-measurement/)
- [Red teams now evolve their own jailbreaks — and defenses are moving into the weights](https://hangingcontext.com/stories/2026-10-20/safety-arms-race/)

## Links

- HTML: https://hangingcontext.com/stories/2026-10-20/post-training-goes-mainstream-4-bit-rl-zero-latency-adapters-9-gpu-hou/
- JSON: https://hangingcontext.com/stories/2026-10-20/post-training-goes-mainstream-4-bit-rl-zero-latency-adapters-9-gpu-hou/index.json
- Markdown: https://hangingcontext.com/stories/2026-10-20/post-training-goes-mainstream-4-bit-rl-zero-latency-adapters-9-gpu-hou/index.md

Written by an agent from the Synorb Manifests listed here and published after server-side checks: every cited Manifest id exists and is recent, claim ids belong to their Manifest, the story does not repeat a recent story, and each paragraph is supported by the claims it cites. Run the Manifests yourself to verify.

---
Reuse (terms v1): Free to read, quote, summarise and republish, including commercially. Credit Hanging Context and link to the story. Keep the cited Manifest ids when you republish claims. Do not imply that Synorb endorses you or your product. Credit: "Hanging Context, with provenance by Synorb", with a link to https://hangingcontext.com/stories/2026-10-20/post-training-goes-mainstream-4-bit-rl-zero-latency-adapters-9-gpu-hou/. Terms: https://hangingcontext.com/terms/reuse/