DeepSeek Claims 42.3 on Terminal-Bench 4.0 for V4.1-Flash-1017; Vals AI Measured 31.8

DeepSeek's V4.1-Flash-1017 scores 42.3 on Terminal-Bench 4.0 in the vendor's own table, up from 31.2 for the September V4.1-Flash. Vals AI, which ran the model itself, recorded 31.82%.

The weights went up on Hugging Face on Oct. 17 under the MIT licence, and are not gated. The model card calls the release "a minor update" to V4.1-Flash. TechNode reported it on Oct. 19 and listed no benchmarks, price or parameter count.

Two Terminal-Bench 4.0 numbers

The card's table says DeepSeek used the Claude Code harness for Terminal-Bench 4.0. It lists 42.3 for the new model, 31.2 for the previous Flash, 39.9 for GPT-5.6 Sol, 37.9 for GLM-5.3 and 51.8 for Opus 5. All of these are vendor-supplied.

Vals AI lists 31.82% (plus or minus 2.62) on the same benchmark, ranked 17th of 52 models. The Vals page does not say which harness it used, so the 10.5-point gap cannot be attributed to the model alone. It is also close to the 31.2 DeepSeek reports for the model it replaced.

1. V4.1-Flash-1017, DeepSeek's figure42.3 points
2. V4.1-Flash-1017, Vals AI31.82 points
3. V4.1-Flash, DeepSeek's figure31.2 points

Source: DeepSeek model card (vendor-supplied); Vals AI model page; accessed 2026-10-21

On other rows the picture is mixed. The card has the new model at 77.0 on DeepSWE v1.1 against 74.0 for Opus 5, using the Mini-SWE-V2 harness. On Humanity's Last Exam without tools it lists 42.6 against 56.3 for Opus 5.

Vals puts the model 8th of 23 on its Vals Index at 61.26%, with a cost of $0.657 per test and a run time of 39 minutes 55 seconds.

What the card says about the model itself

The card gives no parameter count. The config file shows 40 text layers, a hidden size of 5,120, 384 routed experts plus one shared expert, and 6 experts per token. Weights are FP8, with the experts stored in FP4. Context is 1,048,576 tokens, and the card recommends a max_tokens setting of at least 256K.

Hugging Face's metadata counts 763,205,315,794 safetensors entries, but 557 billion of them are 8-bit integer entries and the config says the experts are stored in FP4, so the figure is not a reliable parameter total.

The release has no Jinja chat template. DeepSeek ships a Python reference encoder in the repository and a separate Rust toolkit, deepseek-recipe, for production prompt handling.

API pricing and what changed for existing users

The pricing page lists the model as deepseek-flash, with 1M context and a 384K output cap. Prices are halved off-peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays.

Price per 1M tokens	Off-peak	Peak
Input, cache hit	$0.003	$0.006
Input, cache miss	$0.15	$0.30
Output	$0.60	$1.20

The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired, according to the page. Requests to them are served by V4.1-Flash-1017 and billed at the Flash price, so an application that pinned either name is now running the new weights without a change on its side.

The larger deepseek-v4-pro, version 0813, is listed at $0.66 for cache-miss input and $1.98 for output off-peak.

Why the comparison matters

DeepSeek says the update cuts task time and output length and is "advancing the performance–cost Pareto frontier across multiple agent harnesses". The card publishes no latency or token-count figures to check that. Vals's 39 minutes 55 seconds is a total for its own test set and has no earlier DeepSeek run beside it on that page.

For another agent benchmark the site has covered, see the RuneBench report. For another open release aimed at inference hardware, see Google's TorchTPU release.

The technical report linked from the V4.1-Flash card is the document to check next for the training and harness details this release leaves out.