Skip to content

Open models

DeepSeek’s minor Flash update beats Opus 5 on DeepSWE

5 min read

Our AI analyst checked this story itself: it ran, measured and compared against the originals. What we checked

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: absolutegeeks.com

DeepSeek on October 17 put new weights for its V4.1-Flash model on Hugging Face as DeepSeek-V4.1-Flash-1017. DeepSeek’s model card scores the build at 77.0 on DeepSWE v1.1, up from 74.2 for the previous Flash and ahead of Opus 5 at 74.0. Only the weights changed. A team already running V4.1-Flash can swap in the new shards and keep its config, tokenizer and prompt code, and API calls to deepseek-flash already land on 1017.

The card also says 1017 finishes tasks faster and with shorter output, which DeepSeek says makes each task cheaper. The model is live in DeepSeek Chat as well, and the weights are under the MIT License.

What the API serves now

DeepSeek’s API pricing page names DeepSeek-V4.1-Flash-1017 as the model behind deepseek-flash. Per 1M tokens, peak rates are $0.006 for cache hits, $0.30 for cache misses and $1.20 for output. Off-peak is half that: $0.003, $0.15 and $0.60. Off-peak means every hour outside 01:00-04:00 and 06:00-10:00 UTC on weekdays. The context window is 1M tokens, and output tops out at 384K.

The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are, in the pricing page’s words, “served by the DeepSeek-V4.1-Flash-1017 model and billed at the Flash price”. Code still pinned to either old name is therefore running 1017 too.

For agents, the cache-hit line matters most. DeepSeek said when it introduced V4.1-Flash on September 10 that cache-hit charges often make up a large share of agent costs. It also said the model’s KV cache (the stored attention state of earlier tokens) needs a quarter of the HBM and an eighth of the SSD storage the previous generation used.

Against Opus 5 and GLM-5.3

DeepSeek calls 1017 a minor update, yet its Terminal-Bench 4.0 score jumped 11.1 points, from 31.2 to 42.3, measured in the Claude Code harness. ProgramBench went from 20.3 to 29.1, Apex Bench v1.1 from 64.6 to 72.1 and HLE without tools from 36.8 to 42.6. The DeepSWE scores come from the Mini-SWE-V2 harness.

Opus 5 still beats the new Flash on 7 of the card’s 11 rows. Among them: Terminal-Bench 4.0 at 51.8, NL2Repo-Bench at 75.3 against 68.2, and HLE without tools at 56.3. Flash-1017 is ahead of Opus 5 on DeepSWE, Automation-Bench v1.0.6 (60.5 vs 50.3), Agents’ Last Exam (35.4 vs 28.6) and Apex Bench (72.1 vs 65.8). It also beats GLM-5.3 on all eight rows where GLM-5.3 has a score. On DeepSWE it leads Kimi K3 at 67.5 and GPT-5.6 Sol at 73.0 as well. Every figure in the table comes from DeepSeek.

Running the weights yourself

V4.1-Flash is a 552B-parameter mixture-of-experts model with native image understanding. It uses 8B active parameters for input and 16B for output. The config stores weights in FP8, with the experts in FP4. The 1017 shards total 510,296,708,312 bytes, the same as the previous build. The card recommends temperature 1.0, top_p 1.0 and max_tokens of at least 256K, with a 1M-token context window. The inference folder covers weight conversion and running the model locally.

The 1017 repo ships without a Jinja chat template, the tech-report PDF and the evaluation folder. The base V4.1-Flash repo has had a 15,348-byte template since October 1, ported from encoding.py, the Python reference encoder. That port came before an October 16 fix to encoding.py titled “preserve tool image order and current-turn thinking”, and the copy of encoding.py in 1017 already includes that fix. For production, the card points to deepseek-recipe, a set of Rust libraries with Python bindings. They turn Messages, Chat Completions and Responses API requests into V4 and V4.1 prompts or token IDs and parse the output back. Inference, tool execution and HTTP stay with the caller.

Our analysis

What we checked

On October 18 we compared the 1017 repository file by file with the base DeepSeek-V4.1-Flash repo, read DeepSeek’s API pricing page, and ran deepseek-recipe 0.1.1 against the card’s reference encoder.

How we check stories

Apart from the card text, only the weights differ between the two repos. The card lists no price and the 1017 repo has no chat template, yet DeepSeek publishes both: the price on its pricing page, the template in the base repo. The Rust toolkit the card recommends matches the reference encoder on tools and plain chat. In thinking mode it renders earlier turns differently.

  • Of the 83 files both repos share, 33 are byte-identical, including the config, tokenizer files, the MIT license and everything under encoding/ and inference/. The other 50 are the README, .gitattributes and the 48 weight shards.
  • The card’s test suite passes 55 of 55. The recipe’s output matches the reference byte for byte on one fixture covering tools, tool calls, reasoning effort and thinking mode, and in chat mode on a second.
  • The reference’s "task": "action" field renders as <|action|>Search, and a Chat Completions request cannot express it.
  • deepseek-recipe was last pushed on September 10, before the October 16 encoder fix. We have not measured whether it reproduces that fix.
  • In thinking mode, the recipe keeps the earlier assistant turn’s reasoning (<think>…</think>Hi there!) where the reference drops it (</think>Hi there!). For history turns without reasoning, the recipe emits <think></think> where the reference emits </think>. We tested Chat Completions requests with default conversion options and did not test how this changes the model’s answers.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.