Checkpoint
also called Weights snapshot, Model snapshot
A saved copy of a model’s weights at one point in training or post-training.
A complete saved copy of a model’s weights at a particular moment — after a training run, after a round of post-training, or after fine-tuning. The checkpoint is the thing you actually serve, download or evaluate: the architecture may stay the same while the weights change, so two checkpoints of the “same” model can behave differently.
The practical trap is the name. Vendors serve updated checkpoints under the same model alias, so the same alias in September and in October can be different weights with different benchmark scores, latency and output length. When you compare results over time — or reproduce someone else’s — the dated checkpoint matters more than the model name. Pin the checkpoint or the dated release when you can; when you cannot, record the date you ran the eval.
DeepSeek’s October 2026 refresh is the textbook case: V4.1-Flash-1017 shipped October 17 under the existing deepseek-flash alias, with vendor-reported gains — 77.0 on DeepSWE v1.1, up from 74.2 — and shorter outputs. The weights are MIT-licensed on Hugging Face, so self-hosters can pin the exact dated release; API callers get whatever the alias currently serves. The same pattern ran three months earlier, when the 0731 checkpoint superseded the V4-Flash preview with architecture unchanged and gains from re-post-training.
The benchmark that moved
A team’s eval suite shows a sudden jump on a model they did not change.
Before celebrating — or filing a bug — check whether the provider refreshed the checkpoint behind the alias. If the vendor publishes dated releases, re-run the suite against the pinned old checkpoint to separate a real capability change from a serving change.
Quoting a benchmark for a model name without the checkpoint date. The number may belong to weights nobody serves any more.
- reportDeepSeek refreshed Flash weights and its live API under the same model alias ↗
The October 17, 2026 refresh: new V4.1-Flash-1017 weights served under the unchanged deepseek-flash alias, making the dated checkpoint the thing that matters for comparisons.
- reportDeepSeek Just Released V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API ↗
Release details for the 1017 checkpoint, including the vendor-reported 77.0 on DeepSWE v1.1 and the MIT-licensed weights on Hugging Face. Benchmark figures are the vendor’s, not independently reproduced here.
- reportDeepSeek Upgrades DeepSeek-V4-Flash-0731 with Major Agentic and Coding Gains ↗
The earlier precedent: the July 31, 2026 checkpoint superseded the V4-Flash preview with architecture unchanged, gains from re-post-training.
Weights · Parameters
The numbers inside the model that training moved.
Fine-tuning
Extra training on your own examples to change how the model behaves.
Pretraining
The first, most expensive pass: predict text at internet scale.
Evaluation · Evals
How you know the system is actually working.
Model Provider
The company or platform you call to run inference.