1. Home
2. News
3. Open weights

DeepSeek V4.1 Flash 1017 ranks second among open-weight models, one point behind MiMo-V2.6-Pro

It gained six points on Artificial Analysis's index in five weeks, while NVIDIA squeezed a 74B model into 24GB for laptops.

By Think Facility · October 19, 2026 · DeepSeek Open weights Save

DeepSeek put out an update to V4.1 Flash on October 17, 2026, called DeepSeek-V4.1-Flash-1017 after its date. The weights are on Hugging Face under the MIT license, and DeepSeek says the new checkpoint is already live on its website, app and API. Anyone calling the API by the name deepseek-flash now gets it.

It's close to the top of the open-weight field without being first. Artificial Analysis, which runs its own tests on every model it lists, scores 1017 at 45 on its Intelligence Index, one point behind Xiaomi's MiMo-V2.6-Pro at 46 and level with GLM-5.3.

DeepSeek V4.1 Flash 1017

45 on the Intelligence Index

Highest-ranked open weights model

MiMo-V2.6-Pro, 46

Earlier V4.1 Flash (max)

39

Nemotron 3.5 Bolt 74B-A9B

74B total, 9B active, 24GB

What changed since September

When we wrote about the first V4.1 Flash on September 15, the story was one independent test where it had edged past GPT-6 Astra. On the same index today, the original V4.1 Flash at max effort reads 39. The 1017 checkpoint reads 45, a six-point gain for what DeepSeek's card calls a minor update.

DeepSeek's own table shows where it came from. Terminal-Bench 4.0, a test of working in a command line, went from 31.2 to 42.3. DeepSWE v1.1, its software engineering test, went from 74.2 to 77.0, and Humanity's Last Exam without tools from 36.8 to 42.6. Opus 5 is still well ahead in the same table on Terminal-Bench (51.8) and Humanity's Last Exam (56.3).

The card also says 1017 finishes tasks faster with shorter outputs. Artificial Analysis still counts it as very verbose (210M output tokens to get through the index, where the median is 140M), and it runs at 224.7 output tokens a second.

NVIDIA's Nemotron 3.5 Bolt fits in 24GB

NVIDIA released Nemotron 3.5 Bolt 74B-A9B 24GB on Hugging Face on October 15. It's a mixture of experts with 74B parameters, 9B of them active, cut down from NVIDIA's 120B Nemotron 3.5 Super VL. To get it into 24GB, NVIDIA stored nearly all the expert weights at about two bits each (2.0625 bits per element, to be exact) and trained it further so the shrinking cost less accuracy.

NVIDIA's table puts Bolt ahead of Qwen3.6 35B-A3B and Gemma 4 26B-A4B on most rows: BrowseComp, a web-research test, is 66.75 against 48.74 and 26.30. Artificial Analysis lists Bolt at 21, Qwen3.6 35B A3B at 18 and Gemma 4 26B A4B at 17, which is the same order with smaller gaps. It isn't faster than those two, though. NVIDIA's own speed test gives Bolt 83.99 tokens a second on coding against 96.63 and 100.56.

What a 24GB model means on a desk

NVIDIA built Bolt for local agents on its RTX Spark and DGX Spark machines, and its blog ran the speed tests on a Surface Laptop Ultra with 128GB. The card lists a GeForce RTX 5090 among the test hardware too. It doesn't name a 24GB graphics card, and its vLLM recipe targets A100, H100 and Blackwell GPUs, so "fits in 24GB" describes the packed model on disk and in memory. The card gives no result for a 24GB card.

The licenses matter if you'd build on either. DeepSeek's weights are under MIT. Bolt comes under OpenMDW-1.1, and NVIDIA's card says it's ready for commercial or non-commercial use, so a small company could run either one on its own hardware.

DeepSeek's announcement says billing for its Search and Fetch tools inside Claude Code, Codex and its own harness starts on October 31, 2026, at 16:00 UTC.