DeepSeek has just released DeepSeek-V4.1-Flash-1017, a new checkpoint of its V4.1-Flash model. It is live on the DeepSeek website, app and API. DeepSeek claims stronger scores, faster task completion and shorter outputs than the September V4.1-Flash release. The company calls it a Pareto improvement in performance and cost. Alongside it, DeepSeek opened its Search and Fetch APIs to every developer. For agent builders, that means a cheaper coding model plus a first-party web layer behind one API key.
TL;aDR
- Size: 552B backbone parameters plus 196B Engram parameters. It activates 8B parameters per token in prefill and 16B in decode. Context is 1M tokens, with up to 384K output tokens.
- Runs on: DeepSeek API as
deepseek-flash. September V4.1-Flash weights (763B parameters, FP8) are MIT-licensed on Hugging Face. Hardware requirements and 1017 weights: Not disclosed. - Performance: Beats the previous V4.1-Flash on all 11 benchmarks DeepSeek reported, by 2.0 to 11.1 points.
- Best: 77.0 on DeepSWE v1.1, above Claude Opus 5 (74.0) in DeepSeek’s table.
- Worst: 42.6 on HLE without tools, versus 56.3 for Claude Opus 5.
What is DeepSeek-V4.1-Flash-1017?
DeepSeek-V4.1-Flash-1017 is an updated post-training checkpoint of DeepSeek-V4.1-Flash, served through the deepseek-flash model name. The architecture details below come from the V4.1-Flash technical report.
DeepSeek team states that DeepSeek-V4.1-Flash-1017 uses reinforcement learning with self-feedback for end-to-end project work. Its examples run from game creation to 3D modeling. It is tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. DeepSeek recommends DeepSeek Harness v0.2.1-rc.1 or later.
How does the V4.1-Flash architecture work?
V4.1-Flash is a multimodal Mixture-of-Experts model with 40 Transformer layers. They split into a 20-layer causal encoder and a 20-layer decoder.
- Causal Encoder-Decoder (CED): Decoder global KV is projected from the final encoder hidden states. This nearly halves prefill compute.
- Compressed Sparse Attention 2 (CSA2): Layers run in Full, Reindex or Reuse mode. They share main KV and indexer keys across layers.
- FP4 main KV cache: Global KV drops to 890 bytes per token, about 1/4 of V4-Flash.
- SWA Bounded Replay: Persistent KV cache shrinks to roughly 1/8 of V4-Flash.
- Reasoning effort: A 1 to 100 scalar. The API tiers low, high and max map to 50, 75 and 100.
DeepSeek-V4.1-Flash-1017 benchmarks
All scores below are DeepSeek’s own figures. DeepSeek ran Terminal-Bench 4.0 in Claude Code and DeepSWE in Mini-SWE-V2. ZeroBench-main uses Pass@5.
| Benchmark | V4.1-Flash-1017 | V4.1-Flash | Kimi K3 | GLM 5.3 | GPT 5.6-Sol | Claude Opus 5 |
|---|---|---|---|---|---|---|
| DeepSWE v1.1 | 77.0 | 74.2 | 67.5 | 66.9 | 73.0 | 74.0 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 12.6 | 37.9 | 39.9 | 51.8 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 46.7 | 48.8 | 45.8 | 50.3 |
| Agents’ Last Exam | 35.4 | 31.8 | 27.6 | 28.5 | 26.7 | 28.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | 47.8 | 56.6 | 54.3 | 65.8 |
| ProgramBench | 29.1 | 20.3 | 17.5 | 19.0 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 58.0 | 58.0 | 56.8 | 75.3 |
| HLE (no tools) | 42.6 / 44.0* | 36.8 / 39.1* | 43.5 | 42.0* | 44.5 | 56.3 |
| BabyVision (w/ tools) | 92.5 | 89.6 | 85.7 | Not disclosed | 88.9 | 94.1 |
*Text-only HLE subset.
The 1017 checkpoint leads on DeepSWE, Automation-Bench, Agents’ Last Exam and Apex Bench. Claude Opus 5 still leads on Terminal-Bench 4.0, ProgramBench, NL2Repo and HLE.
‘;});
document.getElementById(‘dsx-bars’).innerHTML=h;document.getElementById(‘dsx-bnote’).textContent=b.n;
requestAnimationFrame(function(){requestAnimationFrame(function(){R.querySelectorAll(‘#dsx-bars .bar’).forEach(function(e){e.style.width=e.dataset.w+’%’;});});});
setTimeout(size,50);}
R.querySelectorAll(‘#dsx-bsel .chip’).forEach(function(c){c.addEventListener(‘click’,function(){R.querySelectorAll(‘#dsx-bsel .chip’).forEach(function(x){x.classList.remove(‘on’)});c.classList.add(‘on’);draw(+c.dataset.b);});});
var S=[
{t:’1. Causal Encoder-Decoder (CED)’,p:’40 layers split into a 20-layer encoder and a 20-layer decoder. Decoder global KV is projected from the final encoder states, so prefill activates 8B params and decode 16B.’,
s:’‘},
{t:’2. CSA2: Full, Reindex, Reuse’,p:’Each attention layer runs in one static mode. Full computes KV and Top-K indices. Reindex reuses KV but rescores. Reuse borrows both, cutting cache and indexing work.’,
s:’‘},
{t:’3. Smaller KV cache’,p:’FP4 main KV plus cross-layer sharing cut global KV to 890 bytes per token, about 1/4 of V4-Flash. SWA Bounded Replay cuts persistent cache to about 1/8.’,
s:’‘},
{t:’4. Controllable reasoning effort’,p:’Effort is a 1 to 100 scalar learned in RL. The API exposes low, high and max tiers, mapped to 50, 75 and 100. Higher effort means longer, more accurate traces.’,
s:’‘}
];
var st=0,dots=R.querySelectorAll(‘.dot’);
function show(i){st=i;var x=S[i];document.getElementById(‘dsx-stage’).innerHTML=’
‘+x.t+’
‘+x.p+’
‘+x.s;
dots.forEach(function(d,k){d.classList.toggle(‘on’,k<=i)});document.getElementById(‘dsx-prev’).disabled=i===0;document.getElementById(‘dsx-next’).disabled=i===S.length-1;setTimeout(size,50);}
document.getElementById(‘dsx-prev’).onclick=function(){if(st>0)show(st-1)};
document.getElementById(‘dsx-next’).onclick=function(){if(st
