🐋 DeepSeek-V4.1-Flash-1017

552B MoE,
stronger on agents.

A streaming demo of the 1017 checkpoint — the update that lifts Terminal-Bench 4.0 from 31.2 → 42.3 and Apex Bench from 64.6 → 72.1 while cutting output length. Weights stay on the inference providers; this page is the client.

552B total · 8B/16B active 1M context FP8 · 510 GB MIT
Try it I want the weights

Chat

not connected
Connect an access token required

A static Space has no backend, so there is nowhere to keep a secret. Pasted tokens are held in this tab's sessionStorage and sent only to router.huggingface.co; nothing is proxied through this Space's owner, and inference is billed to your account. Create one at settings/tokens (read scope is enough) — or run this Space's owner-funded version with PRO hardware.

No token in this session.

No messages yet. Ask for an irrationality proof, a palindrome function, or the causes of WWI in three sentences.

Reasoning stream 0 chars 

        


        

          
          

            
            
            
          

        

        

          This checkpoint reasons before answering — expect several seconds of
          reasoning stream before the first visible token. A UI that only renders
          content looks frozen.
        

      


      

        

          
1017 vs V4.1-Flash

          from the model card · 1017 in blue, previous release in grey
        

        

        

          Biggest jumps: Terminal-Bench 4.0 +11.1, ProgramBench
          +8.8, Apex Bench +7.5, HLE w/o tools
          +5.8. Every benchmark below moves up, none regress.
        

      


      

        

          
Running the actual weights

          
        

        

          No Space tier can host this model: 552B params, ~510 GB in 48
          FP8 shards, validated upstream on 4× H200 (141 GB) on one NVLink
          domain, ~112 cores, ~1180 GB host RAM. Cold start is ~28 minutes.
        

        
docker run --gpus all --shm-size 32g -p 30000:30000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --ipc=host --env HF_TOKEN=$HF_TOKEN lmsysorg/sglang:dev-dsv41 \
  sglang serve \
    --model-path deepseek-ai/DeepSeek-V4.1-Flash-1017 \
    --tp-size 4 --ep-size 4 \
    --context-length 262144 --mem-fraction-static 0.85 \
    --reasoning-parser deepseek-v41 --tool-call-parser deepseekv41 \
    --trust-remote-code

        
      


      

        
About this checkpoint

        

          Official: deepseek-ai/DeepSeek-V4.1-Flash-1017
          (131k downloads, 3.9k likes). A separate upload,
          oakshznan/DeepSeek-V4.1-Flash-1017, appeared on 2026-10-30 with the
          same 510,296,920,807-byte storage and no provider mapping — this demo targets the
          official repo. If you pull the other one, diff shard checksums first.
        

        

          This release ships encoding/encoding.py instead of a Jinja chat
          template, so the page sends plain chat messages and lets each provider apply its
          own encoding.