Sulat.com / Writing on AIThursday, November 19, 2026
Artificial Intelligence/18 November 2026

Together AI opens DeepSeek V4.1 Flash 1017 to post-training, but its docs don’t list the model yet

Together AI says you can now post-train DeepSeek V4.1 Flash 1017 with SFT or RL on your own data, keeping the loop in your Python. Its post-training page and docs catalog don't list the model yet.

Glossy blue 3D DeepSeek whale ringed by a glowing training loop with tuning dials, above a row of GPU chips.

At a glance

  • Together AI said on X early on 18 November UTC that DeepSeek V4.1 Flash 1017 is available for post-training, with SFT or RL on your own data, rewards and environments and the loop in your own Python.
  • One RL step is five calls: sample(), your_reward(), forward_backward(), optim_step() and weights_sync(). SFT skips sample() and weights_sync(), and the reward and sandbox are your own code.
  • DeepSeek released the 1017 update on 17 October under the MIT licence with a 1M-token context, and its own table has DeepSWE v1.1 at 77.0 against 74.2 and Terminal-Bench 4.0 at 42.3 against 31.2.
  • Together's post-training page and supported models catalog don't list the model yet, so LoRA or full-weight support, context limits and an hourly rate aren't published.
  • Post-training bills per GPU-minute only while a resource is READY, and a resource keeps billing after a script exits unless something stops it.

Together AI says you can now post-train DeepSeek V4.1 Flash 1017 on its platform. The announcement went up on X early on 18 November UTC, and it’s short: SFT or RL, your own data, your own rewards and environments, and a training loop that stays in your Python.

The picture in the post is a screenshot of Together’s console. It shows the Create model resource dialog with deepseek-ai/DeepSeek-V4.1-Flash-1017 highlighted in the base model list.

You write the loop, Together runs the GPUs

What does Together actually do here? Less than you might guess, and that’s the point. On its post-training page, you write the recipe: data, rewards, environments, the loss, the loop. Together handles the trainer and, when you’re doing RL, the rollout engine and the weight sync between the two. Pick GRPO, PPO or a loss you’ve written yourself. Train a LoRA adapter, or full weights if the model allows it.

One RL step is five calls: sample(), your_reward(), forward_backward(), optim_step() and weights_sync(). SFT skips sample() and weights_sync(). Your reward is plain code, so it’ll work with whatever sandbox you already run, and for multi-turn agents your script drives the episode while Together generates each turn.

Split diagram: your Python holds the loop and your_reward(); Together runs sample, forward_backward, optim_step and weights_sync.
You write the loop and the reward; Together runs the other four calls.

The model underneath

DeepSeek shipped the model itself on 17 October. It called 1017 a minor update to V4.1 Flash, put the weights out under the MIT licence and kept the 1M-token context. The model can read images. Together’s docs, though, say post-training only trains and samples on text, even for models that can.

DeepSeek’s own table makes the update look big for agentic coding. DeepSWE v1.1 climbs from 74.2 to 77.0. Terminal-Bench 4.0 jumps from 31.2 to 42.3. Claude Opus 5 is still ahead on Terminal-Bench 4.0 (51.8) and on ProgramBench (37.0 against 29.1). These are all DeepSeek’s numbers, run on the Claude Code harness for Terminal-Bench and Mini-SWE-V2 for DeepSWE.

Paired bars: DeepSWE 74.2 to 77.0, Terminal-Bench 4.0 31.2 to 42.3, ProgramBench 20.3 to 29.1, all DeepSeek-reported.
DeepSeek’s own table has 1017 ahead of V4.1 Flash on all three coding benchmarks.

What Together hasn’t published yet

Together’s own pages don’t list the model yet. The post-training page shows GLM-5.3, Kimi K3, Qwen3.8 27B and Qwen3.6 35B A3B, plus a “coming soon” tile. The supported models catalog has nine entries, none of them DeepSeek. So there’s no word yet on whether 1017 trains as LoRA only or with full weights, how long a sequence the trainer takes, or what an hour costs. The docs add that everything runs on B200 GPUs for now.

The billing is simple. You pay per GPU-minute, and only while a resource is READY. Queueing, starting up and shutting down are free, and so are sessions, which means several LoRA experiments on one resource cost the same per minute as a single one. You can ask for an hourly estimate with estimate_cost before creating anything. In the pricing docs‘ worked example, Qwen3.8 27B with LoRA and one generator replica comes to US$88 an hour at US$11 per GPU-hour, and a new resource needs credit for four hours of it.

Watch for one trap. A resource keeps billing after your script exits unless something stops it. The cookbook recipes clean up after themselves, and an idle resource stops on its own after an hour.

Four states: queued, starting and shutting down are free; READY is billed per GPU-minute. Idle resources stop after an hour.
The meter runs only while the resource is READY, and it keeps running until something stops it.

Sources


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.