Ember-1

	Ember-1(fireworks.ai)
	585 points by gmays 16 days ago \| 249 comments
	tl;dr: Fireworks Research released Ember-1, a Kimi K3-derived model trained to cut reasoning tokens by ~40% while matching K3's quality on benchmarks, live A/B tests, and internal coding workloads. It sets a Pareto frontier on cost-vs-quality across benchmarks like SWE-bench Verified, Terminal Bench, and Doximity's Bedside Bench, beating GPT-5/6, Claude Opus 5, and GLM 5.3 on cost/task. It's available as a two-week research preview on Fireworks Serverless, with training support for custom fine-tuning.
	HN Discussion: ↓Fireworks is unethically closing weights derived from open Chinese models like Kimi K3 ↓The pricing doesn't justify the token reduction since per-token cost is doubled ↓Skepticism about Fireworks shifting from neutral inference provider to model developer competing with the OSS models it serves •Enthusiasm for the broader golden age of accessible model training and fine-tuning ↑Agreement that reducing over-thinking in reasoning models is genuinely valuable