DeepSeek V4.1 Flash 1017 lifts Terminal Bench from 31.2 to 42.3

DeepSeek has released the V4.1 Flash 1017 checkpoint and deployed it to its website, app and API under the existing deepseek-flash identifier. The company also published the weights under the MIT license. The 1017 label marks the October 17 release date rather than a new model family.

The October checkpoint advances most on agent benchmarks

DeepSeek's tables show Terminal Bench 4.0 rising from 31.2 to 42.3. DeepSWE v1.1 moves from 74.2 to 77.0 and ProgramBench from 20.3 to 29.1. The vendor also says the model finishes tasks faster with shorter outputs while API prices remain unchanged.

V4.1 Flash 1017 still accepts text and images. That capability is not new to the October checkpoint. Native vision arrived with the base V4.1 Flash release in September; the new part is the updated weights and reported performance gain.

Shorter outputs may matter more than a higher score

Agent tasks do not incur the cost of one prompt alone. A model repeatedly reads state, calls tools and generates subsequent steps. Reaching the answer with fewer tokens and less wall-clock time can lower the cost of the whole run even when the price per million tokens stays fixed.

Open weights also allow private deployment and modification. That freedom comes with a bill: this is a very large MoE checkpoint, not a model for one ordinary GPU. For many companies, the hosted API will remain cheaper than operating a server cluster.

The GLM 5.3 comparison still rests on uneven evidence

Nathan Lambert suggested the model could beat GLM 5.3 while including vision. That remains an estimate, not an independently established result. Agent benchmarks are sensitive to the harness, enabled tools, reasoning effort and task budget.

A vendor table also cannot tell a team how the model behaves on its repository or interface. A higher score earns the checkpoint a place in a test shortlist, not an automatic route into production.

In-house tasks will show whether the checkpoint saves time and money

Independent evals should run DeepSeek, GLM and other open-weight models with the same harness and budget. Teams need to measure end-to-end task success, token use, runtime and human intervention.

Operation outside DeepSeek's API matters too. Support in vLLM and SGLang, memory requirements and tool-use stability will determine whether the open checkpoint broadens deployment options or mainly powers an inexpensive hosted service.

Lilith's verdict

DeepSeek has brought a faster runner, but DeepSeek still holds most of the stopwatches. Before handing it the production relay, time it on your own track.