Together AI: fine-tune 18 new open models, MoE LoRA, 30-70% off
MoE adapters recall 89% of tuned knowledge vs 15% before, and runs refund unused compute. Plus: OpenAI retires Codex-Spark and reworks GPT-6 Astra prompts.
Copy markdown
18 models join the fine-tune menu
GLM 5.3, Kimi K2.7-Code, DeepSeek-V4-Flash, Qwen 3.x and Gemma 4 are now fine-tunable on Together — 18 open-weight additions in one drop. You can train the coding and agent models you already run in production instead of settling for a stock checkpoint.
Expert LoRA finally fits MoE
New adapters target mixture-of-experts layers, not just attention, lifting knowledge recall to 89% from 15%. For teaching a modern MoE model your domain, that's the gap between a fine-tune that sticks and one that forgets what you taught it.
Training just got 30-70% cheaper
LoRA SFT on gpt-oss-120b drops from $5 to $2.50 per million tokens; Qwen 3.5-9B falls to $0.34 from $0.48. Early stopping halts a run at the validation-loss plateau and auto-refunds the compute you didn't burn.
See the run before it bills you
Live metrics, job status and artifacts now stream to the dashboard and API, dataset previews tokenize your rows with the exact chat template, and pre-flight validation rejects a bad file before the job starts charging.
Elsewhere: OpenAI pulls Codex-Spark next week
OpenAI's Codex lead says GPT-5.3-Codex-Spark — the Pro-only fast model — reportedly goes offline next week on falling usage, with a quicker replacement promised. If your Codex config pins Spark, switch before it disappears.
Elsewhere: Rework your prompts for GPT-6 Astra
OpenAI tells devs to slim down Codex prompts and AGENTS.md for Astra: make skill triggers specific, load guidance only when relevant, and define what 'done' means. Heavy scaffolding now backfires — there's even a slop-word blocklist.