ModelsArchitectures & capability
DeepSeek Releases V4.1-Flash-1017 Open Weights Under MIT License and Launches Search API
DeepSeek has published MIT-licensed weights for V4.1-Flash-1017, an in-place upgrade to its deepseek-flash API model at unchanged pricing, pairing higher coding-agent benchmark scores and lower per-task costs with a new first-party Search and Fetch API.
From the huggingface.co article of October 17, 2026 and cited sources: Released on October 17, the 1017 checkpoint replaced the model behind the deepseek-flash API name without a version pin and kept the same pricing. [1] [4] [7]
SOURCES[1] deepseek-ai/DeepSeek-V4.1-Flash-1017 · Hugging Face huggingface.co[4] DeepSeek API Change Log api-docs.deepseek.com[7] Tech Insider: DeepSeek V4.1-Flash-1017 Hits 77.0, Opens… tech-insider.org
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Released on October 17, the 1017 checkpoint replaced the model behind the deepseek-flash API name without a version pin and kept the same pricing. [1] [4] [7]
- Original source
- huggingface.co article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 17, 2026 article published
- Covered
- Oct 19 edition
- Evidence
- 10 sources across 8 sites
- 1deepseek-ai/DeepSeek-V4.1-Flash-1017 · Hugging Facehuggingface.co
- 2deepseek-ai/DeepSeek-V4.1-Flash · Hugging Facehuggingface.co
- 3Artificial Analysis: DeepSeek V4.1 Flash 1017 (max)artificialanalysis.ai
- 4DeepSeek API Change Logapi-docs.deepseek.com
- 5DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache…arxiv.org
- 6The Decoder: Deepseek V4.1 Flash 1017 beats larger rivals…the-decoder.com
StrongHow it’s measured
Hacker News · 947 likes · 708 comments
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.
Released on October 17, the 1017 checkpoint replaced the model behind the deepseek-flash API name without a version pin and kept the same pricing. Its weights are on Hugging Face under an MIT license. It accepts text and image input and supports a 1M-token context. [1] [4] [7]
The model is a sparse mixture-of-experts encoder-decoder. It uses 16B active parameters during decode and an FP4 KV cache of about 890 bytes per token, which is roughly 4× smaller than V4-Flash's. The 763B and 552B parameter counts listed for it most likely reflect different counting methods. About 196B of conditional 'Engram' memory accounts for most of the difference. [1] [2] [5] [8]
DeepSeek's own scores show DeepSWE v1.1 rising to 77.0 from 74.2 and Terminal-Bench 4.0 rising to 42.3 from 31.2. DeepSeek puts the cost at about $0.20 per task, down from roughly $0.50. Artificial Analysis gives it 45 on its Intelligence Index, second of 121 models in its class. [1] [6] [3]
The weights need DeepSeek's own prompt encoding, because standard chat templates may produce malformed prompts. Teams running the model themselves should budget memory for the full checkpoint, not just the active parameters. Search and Fetch calls are scheduled to cost $0.75 per 1,000 starting October 31. [1] [7]
an open, commercially usable checkpoint with a 1M-token context, a much smaller KV cache and a score of 45 on Artificial Analysis's index.
Read the full assessment
Implication: teams on deepseek-flash should re-run regression tests, because the model changed without a version pin.
Executive brief
DeepSeek's Hugging Face model card lists V4.1-Flash-1017 at 763B parameters. The V4.1-Flash card and Artificial Analysis put it at 552B. The gap seems to come from counting different things, not from a different model. The release is a minor, MIT-licensed checkpoint update that DeepSeek says is both stronger and cheaper per task. Its headline scores are DeepSeek's own, and no one has independently replicated them. The first independent composite score, 45 on Artificial Analysis's index, places it #2 of 121 models in its class. The same evaluator describes it as fast but "very verbose."
What changed and event timeline
(arXiv v1): V4.1-Flash technical report
DeepSeek-AI posted a preprint on KV-cache compression that underpins V4.1-Flash ().
V4-Flash public beta
DeepSeek launched the predecessor line in public beta. Legacy
deepseek-chatanddeepseek-reasonernames were deprecated a week earlier ().V4.1-Flash released
It added native vision, retired V4 Flash and the Vision Exp model (old names temporarily route to V4.1-Flash), and cut API prices ().
1017 checkpoint plus Search & Fetch API
V4.1-Flash was upgraded in place under
deepseek-flashat unchanged pricing, with weights on Hugging Face and ModelScope. A first-party search API launched the same day (;).(scheduled): Search billing begins
Search and Fetch calls cost $0.75 per 1,000. They stay free through DeepSeek Harness, Claude Code and Codex until 16:00 UTC that day ().
Capabilities and access
- Model:
deepseek-ai/DeepSeek-V4.1-Flash-1017, MIT license. Takes text and image input and outputs text, with a 1M-token context and at least 256K recommended output (HF card). - Weights: Safetensors in BF16, F32 and FP8 E4M3. Runs on Transformers, vLLM and SGLang. Recommended settings are temperature 1.0 and top-p 1.0 (HF card).
- Existing integrations get the update with no code changes (Tech Insider).
Read the full section
- Model:
deepseek-ai/DeepSeek-V4.1-Flash-1017, MIT license. Takes text and image input and outputs text, with a 1M-token context and at least 256K recommended output (HF card). - Weights: Safetensors in BF16, F32 and FP8 E4M3. Runs on Transformers, vLLM and SGLang. Recommended settings are temperature 1.0 and top-p 1.0 (HF card).
- API:
deepseek-flash. Existing integrations get the update with no code changes (Tech Insider). - Pricing conflict: Kingy lists $0.15 input / $0.60 output per million tokens off-peak and $0.30 / $1.20 at peak (Kingy AI). OpenRouter shows $0.05 input / $0.60 output.
Technical analysis for researchers and developers
- Architecture: a sparse MoE built as a Causal Encoder-Decoder (CED): a 20-layer causal encoder feeding a 20-layer decoder.
- Memory: "Compressed Sparse Attention 2" (CSA2) with an FP4 KV cache, about 890 bytes per token.
- Parameter count: 552B backbone plus 196B Engram comes to about 748B.
Read the full section
- Architecture: a sparse MoE built as a Causal Encoder-Decoder (CED): a 20-layer causal encoder feeding a 20-layer decoder. It adds 196B parameters of "Engram" conditional memory. Active parameters are 8B during prefill and 16B during decode (V4.1-Flash card).
- Memory: "Compressed Sparse Attention 2" (CSA2) with an FP4 KV cache, about 890 bytes per token. That is roughly 4× smaller than V4-Flash (V4.1-Flash card; arXiv).
- Parameter count: 552B backbone plus 196B Engram comes to about 748B. That suggests the 763B repo label counts all stored tensors (inference). Kingy calls it a different accounting scope.
- Prompt format: prompts use custom encoding (
encoding.pyor the Rustdeepseek-recipe), not Jinja templates. Stock chat templates may produce wrong prompts (HF card). - Training: reinforcement learning with self-feedback (RLSF), according to The Decoder.
Claims and evidence
- DeepSWE v1.1: 77.0, up from 74.2. All three are DeepSeek's own scores (HF card; Kingy).
- Cost per task: about $0.20 at 77.0%, versus about $0.50 for the predecessor. This is DeepSeek's internal figure (The Decoder).
- "Beats Claude Opus 5 and GPT 5.6-Sol." Kingy's reading is that 1017 leads Opus 5 on four benchmarks and trails it on seven (Kingy).
Read the full section
- DeepSWE v1.1: 77.0, up from 74.2. Terminal-Bench 4.0 rose to 42.3 from 31.2, and ProgramBench to 29.1 from 20.3. All three are DeepSeek's own scores (HF card; Kingy).
- Cost per task: about $0.20 at 77.0%, versus about $0.50 for the predecessor. This is DeepSeek's internal figure (The Decoder).
- "Beats Claude Opus 5 and GPT 5.6-Sol." DeepSeek's own comparison. Kingy's reading is that 1017 leads Opus 5 on four benchmarks and trails it on seven (Kingy).
- Independent results:
- Artificial Analysis scores it 45 on its Intelligence Index at $0.26 per index task, running at 227 tokens/s and using 210M output tokens for the full index (Artificial Analysis).
- Kingy ran a 14-submission pilot. DeepSeek passed all 14 and cost $0.016, against $0.058 for GPT-6.1 Sol and $0.065 for Sonnet 5.5. The sample is too small to support a ranking (Kingy).
Context and prior work
- The 1017 checkpoint continues a fast cadence: V4-Flash in July, V4-Pro GA in August with low/high/max thinking levels, then V4.1-Flash in September (DeepSeek Change Log).
- Shorter outputs are DeepSeek's main cost lever. This follows the September KV-compression work (arXiv).
- Third parties have already published inference-engineering work on V4-Flash for AMD gfx90a hardware (arXiv 2609.15627).
Read the full section
- The 1017 checkpoint continues a fast cadence: V4-Flash in July, V4-Pro GA in August with low/high/max thinking levels, then V4.1-Flash in September (DeepSeek Change Log).
- Shorter outputs are DeepSeek's main cost lever. This follows the September KV-compression work (arXiv).
- Third parties have already published inference-engineering work on V4-Flash for AMD gfx90a hardware (arXiv 2609.15627).
- The competitive comparisons named are Claude Opus 5, GPT 5.6-Sol, Kimi K3 and GLM 5.3 (The Decoder).
Limitations, safety and contested findings
- No independent replication of DeepSWE, Terminal-Bench or ProgramBench at release (Tech Insider; The Decoder).
- Shorter-output claim vs. independent measurement: DeepSeek says 1017 produces shorter outputs. Artificial Analysis still rates it "very verbose" (Artificial Analysis).
- Harness sensitivity: the same weights score differently under different harnesses (Kingy).
Read the full section
- No independent replication of DeepSWE, Terminal-Bench or ProgramBench at release (Tech Insider; The Decoder).
- Shorter-output claim vs. independent measurement: DeepSeek says 1017 produces shorter outputs. Artificial Analysis still rates it "very verbose" (Artificial Analysis).
- Harness sensitivity: the same weights score differently under different harnesses (Kingy).
- Undocumented Search API terms: rate limits, billing for failed requests and timeout handling are not specified (Tech Insider).
- Regulatory risk: China is reportedly planning tighter controls on open-weight releases (The Decoder).
Business and practitioner implications
- Silent API change:
deepseek-flashusers are moved to the new checkpoint without a version pin. Re-run regression tests on agent pipelines. - Self-hosting: the MIT license allows commercial use.
- Cost: measure cost per accepted output in your own harness, not list price per token (Kingy). Verbosity can wipe out per-token savings.
Read the full section
- Silent API change:
deepseek-flashusers are moved to the new checkpoint without a version pin. Re-run regression tests on agent pipelines. - Self-hosting: the MIT license allows commercial use. Plan memory for the full checkpoint, not just the 16B active parameters (inference based on the 552B/763B counts).
- Cost: measure cost per accepted output in your own harness, not list price per token (Kingy). Verbosity can wipe out per-token savings.
- Search API: budget for search calls at $0.75 per 1,000 from October 31.