ModelsArchitectures & capability
DeepSeek releases V4.1-Flash-1017 open-weight update, scoring 45 on Artificial Analysis index at $0.26 per task
DeepSeek's V4.1-Flash-1017 checkpoint came out five weeks after V4.1 Flash and uses the same architecture. Its Artificial Analysis Intelligence Index score rose from 39 to 45, and its cost per task fell 47% to $0.26. The weights are MIT-licensed. DeepSeek credits undisclosed self-feedback post-training.
From the Two Minute Papers video of October 28, 2026 and cited sources: Artificial Analysis scored the 1017 checkpoint 45 on its Intelligence Index, up from 39. [2] [3]
SOURCES[2] DEV Community: Artificial Analysis rates 1017 at 45 dev.to[3] Hugging Face: DeepSeek-V4.1-Flash-1017 huggingface.co
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Artificial Analysis scored the 1017 checkpoint 45 on its Intelligence Index, up from 39. [2] [3]
- Original source
- Two Minute Papers video, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- Studies
- arXiv 2507.21931 (Jul 2025) [4] date from arXiv ID
- When
- October 28, 2026 video published
- Covered
- Oct 29 edition
- Evidence
- 11 sources across 11 sites; transcript sourced
RisingHow it’s measured
YouTube video · 103.4K views · 2K likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Artificial Analysis scored the 1017 checkpoint 45 on its Intelligence Index, up from 39. That ranks it second of 121 models tested. Cost per task fell to $0.26 while per-token pricing stayed the same. Vendor-reported model-card scores also rose: Terminal-Bench 4.0 went from 31.2 to 42.3, Humanity's Last Exam (no tools) from 36.8 to 42.6, and DeepSWE v1.1 from 74.2 to 77.0. [2] [3]
The low per-token price ($0.30 input and $1.20 output per million tokens at peak, half that off-peak) can hide high token use. The model produced 210M output tokens on the index, against a median of 140M. In his own unpublished testing, the video's presenter saw it use two to three times more tokens than Claude Opus. A small pilot also found much slower completion than Sonnet on one task. Teams should budget by cost per task, not price per token. [2] [1] [9]
The architecture did not change, so the gains are attributed to post-training that DeepSeek calls self-feedback. DeepSeek has not published the reward design. Some reporting links the method to RLSF, a 2025 technique that uses a model's confidence in its own answers as a reward without human labels. Whether DeepSeek actually uses RLSF is unknown. [1] [3] [4] [10]
The weights are MIT-licensed. The model has a 1M-token context window and can be self-hosted with Transformers, vLLM, SGLang or quantized llama.cpp builds. Running the full model still needs very expensive hardware. The API name stayed the same even though behaviour changed, so practitioners should pin the dated checkpoint and re-run their own evaluations. [3] [7] [1]
Artificial Analysis's index shows higher scores at lower per-task cost within five weeks, with no architecture change. Implication: post-training alone may be advancing open-weight models quickly.
Read the full assessment
Teams that budget by per-token price or rely on unpinned API names may misjudge cost and behaviour.
Executive brief
DeepSeek-V4.1-Flash-1017 is labelled a minor update, but Artificial Analysis scored it 45 on its Intelligence Index, up from 39 for the version released five weeks earlier. Cost per task on that index fell 47%, to $0.26, while the per-token price stayed the same. The weights are open under the MIT licence on Hugging Face. DeepSeek credits "self-feedback" post-training but has published no technical details. In a Two Minute Papers commentary, the presenter says the model is a big step forward. He also warns that the model uses a lot of tokens.
What changed and event timeline
RLSF paper appears
Researchers at Heinrich Heine University Düsseldorf proposed using a model's own answer confidence as a reward signal for post-training.
V4.1 Flash launches
This was DeepSeek's first model on its asymmetric Causal Encoder–Decoder (CED) architecture: a 552B-parameter multimodal MoE priced at $0.15/$0.60 off-peak. It replaced V4 Flash.
V4-Pro folded in
Requests to deepseek-v4-pro started routing to V4.1 Flash, billed at Flash prices.
1017 checkpoint released
DeepSeek describes it as a "Pareto improvement": higher scores with shorter outputs. It went live on DeepSeek's app and API, with MIT-licensed weights.
More detail
Independent index result
Artificial Analysis scored the model 45 (previous version: 39), ranking it #2 of 121 models tested, at $0.26 per task.
Commentary video
Two Minute Papers praised the jump in capability, cautioned about token-heavy behaviour, and speculated on how self-feedback training works.
Capabilities and access
- Model:
deepseek-ai/DeepSeek-V4.1-Flash-1017, served under thedeepseek-flashAPI name (llm-stats). Context is 1M tokens, with generation of at least 256K tokens (Hugging Face). - DeepSeek API pricing: $0.30 input / $1.20 output per million tokens at peak; off-peak is half that (DEV Community).
- Self-hosting: supported through Transformers, vLLM, SGLang and quantized llama.cpp builds (Hugging Face). The video jokes that the hardware costs "one bazillion dollars" (YouTube).
Read the full section
- Model:
deepseek-ai/DeepSeek-V4.1-Flash-1017, served under thedeepseek-flashAPI name (llm-stats). Context is 1M tokens, with generation of at least 256K tokens (Hugging Face). - DeepSeek API pricing: $0.30 input / $1.20 output per million tokens at peak; off-peak is half that (DEV Community). Third-party hosts charge more, for example $1.80 per million output tokens on OpenRouter.
- Self-hosting: supported through Transformers, vLLM, SGLang and quantized llama.cpp builds (Hugging Face). The video jokes that the hardware costs "one bazillion dollars" (YouTube).
Technical analysis for researchers and developers
- Architecture: a sparse MoE on a CED design with a 20-layer encoder and a 20-layer decoder.
- Parameter count is inconsistent: sources give 552B for the backbone but 763B on Hugging Face.
- Training: the 1017 release kept the same architecture (YouTube), so the gains come from post-training. The reward design has not been published.
Read the full section
- Architecture: a sparse MoE on a CED design with a 20-layer encoder and a 20-layer decoder. It has 384 routed experts (6 active per token) plus 1 shared expert. About 8B parameters are active during prefill and 16B during decode, plus a separate 196B-parameter "Engram" memory (Kingy AI).
- Parameter count is inconsistent: sources give 552B for the backbone but 763B on Hugging Face.
- Training: the 1017 release kept the same architecture (YouTube), so the gains come from post-training. DeepSeek names only "self-feedback"; reporting links this to RLSF (BestHub). The reward design has not been published.
- Implementation notes: use DeepSeek Harness 0.2.1-rc.1 or later, and preserve reasoning content across tool calls (Kingy AI). The recommended sampling settings are temperature 1.0 and top_p 1.0.
Claims and evidence
No confidence intervals are reported. The model trails Claude Opus 5 on 7 of 11 benchmark rows (Kingy AI).
Read the full section
- Vendor-reported benchmark gains:
- Humanity's Last Exam (no tools): 42.6, up from 36.8
- Terminal-Bench 4.0: 42.3, up from 31.2
- DeepSWE v1.1: 77.0, up from 74.2
Source: Hugging Face. No confidence intervals are reported. The model trails Claude Opus 5 on 7 of 11 benchmark rows (Kingy AI).
- Vendor-reported cost per task: about $0.20 on DeepSWE, versus about $0.50 for the predecessor (The Decoder).
- Independent: Artificial Analysis gives a score of 45 and $0.26 per task. Output tokens on its index fell from 250M to 210M (DEV Community).
- Conflicting figure: one outlet attributes 74.2 on DeepSWE to the new build (Tech Insider). The model card lists 74.2 as the predecessor's score.
- Anecdotal: the video's claims of "15× cheaper than Opus 5.5" and "2–3× more tokens" come from the presenter's own unpublished testing (YouTube).
Context and prior work
- RLSF ranks multiple reasoning traces by the model's confidence in its final answer, then uses those rankings as synthetic preferences. Whether DeepSeek uses this exact method is unknown.
- The video presents self-critique as an alternative to collecting more training data. Its example is a model that writes a racing game, plays it, and fixes its own bugs (YouTube).
Read the full section
- RLSF ranks multiple reasoning traces by the model's confidence in its final answer, then uses those rankings as synthetic preferences. No human labels are needed (arXiv). Whether DeepSeek uses this exact method is unknown.
- The video presents self-critique as an alternative to collecting more training data. Its example is a model that writes a racing game, plays it, and fixes its own bugs (YouTube). The presenter frames this as his own interpretation of DeepSeek's announcement, not a confirmed description of the method.
Limitations, safety and contested findings
- Verbose output: at 210M output tokens, the model sits well above the index median of 140M (DEV Community).
- Latency: in a small pilot, the model took 27–28 seconds on one task that Sonnet finished in about 3 seconds (Kingy AI).
- Self-judging risk: a model that rewards itself can misjudge its own success (YouTube).
Read the full section
- Verbose output: at 210M output tokens, the model sits well above the index median of 140M (DEV Community). This matches the video's warning to judge cost per task, not per token (YouTube).
- Latency: in a small pilot, the model took 27–28 seconds on one task that Sonnet finished in about 3 seconds (Kingy AI).
- Self-judging risk: a model that rewards itself can misjudge its own success (YouTube).
- Regulatory risk: China reportedly plans tighter export controls on highly capable open-weight models (The Decoder).
Business and practitioner implications
- Budgeting: base budgets on cost per task.
- Pinning versions: behaviour changed substantially within five weeks while the API name stayed the same. Pin the dated checkpoint and re-run your own evaluations.
- Licensing: MIT weights allow on-premise use and distillation.
Read the full section
- Budgeting: base budgets on cost per task. The $0.26 per-task figure assumes peak pricing and Max reasoning effort; off-peak batch runs cost about half (DEV Community).
- Pinning versions: behaviour changed substantially within five weeks while the API name stayed the same. Pin the dated checkpoint and re-run your own evaluations.
- Licensing: MIT weights allow on-premise use and distillation. The video expects distilled versions that run on laptops within a year (YouTube); that is a forecast, not a demonstrated result.
- Choosing a host: API prices vary across hosts, so compare per-task costs before committing to one.
The source trail.
Sources (11)
DeepSeek's New AI Is An Absolute Monster
Transcript retrieved via youtube_auto_captions; language en. Timestamped text, not direct audiovisual review. Source: https://www.youtube.com/watch?v=pryCaGv8wuo. Automatic captions/transcription may contain errors.
www.youtube.com