SYLIN

DeepSeek Flash

by DeepSeek · Model · Language & code models

strong

Best-value open-weights model: near-frontier work for a few dollars, though a step behind.

DeepSeek V4.1 Flash 1017 released October 17 as an update to V4.1 Flash; reviewer calls it the best open model

AccessOpen weightsNeeds~510 GBDownloadabout 510 GB full weights; 1-bit GGUF is 233 GBLicenseMIT

Why we say so

Arena ranks V4.1 Flash second to third among open-weights models on web development and agents, behind Kimi K3.1. AI Search ran it through multi-hour projects for about $6 and found it roughly 80-90% as good as GPT-6 Astra and Opus, with it alone finding the hidden frog in an image test. Much of the benchmark comparison against GLM and Kimi comes from DeepSeek itself, and one HN commenter prefers it over a local Qwen for real work.

Promised Jul 31: Hugging Face weights soon (no date given)

Versions

VersionReleasedRank
DeepSeek V4.1 FlashLATESTSep 9#2 web development (open weights)
DeepSeek V4 FlashJul 31

Evidence

  1. Nov 19
    A swarm of 1,000 agents wrote a 630k-line zero-dependency rule-based chatbot in a week (~8k merged PRs), reported on par with SmolLM-360M
  2. Nov 15
    #2 on Arena webdev among open weights (1659)
  3. Nov 11
    #2 on Arena webdev among open weights (1664)
  4. Oct 23
    Intelligence-versus-cost-per-task chart puts it more cost-efficient than the top GPT and Claude models
  5. Oct 23
    Beats GLM 5.3 and Kimi K3 across the shown benchmarks, including agentic coding and general knowledge
Show all 31
  1. Oct 23
    Benchmark performance roughly on par with GPT Sol, matching the reviewer's experience; below GPT-6 Astra
  2. Oct 23
    Ranks third on KernelBench CUDA, below the latest Claude models but ahead of Gemini 4 Argon, Claude Fable and GPT-6 Astra
  3. Oct 23
    Unreal Engine game with Blender assets: playable but critic scored only 5.5/10 after 16 rounds, about 80% of frontier models
  4. Oct 23
    20-minute Alzheimer's deep-research HTML report judged tied with GPT-6 Astra and much better than Opus 5.5
  5. Oct 23
    Medical deep-research report much better than Opus 5.5, which was lazy and low effort
  6. Oct 19
    Took #1 on Arena agent among open weights
  7. Oct 19
    #3 on Arena vision among open weights (1278)
  8. Oct 18
    Took #1 on Arena webdev among open weights
  9. Sep 27
    MiMo 2.6 is cheaper per task than DeepSeek V4.1 Flash
  10. Sep 18
    552B MoE with encoder-decoder design; open weights (~510 GB), cheap fast API; claims top open-model scores on several leaderboards
  11. Sep 18
    Ranked #1 open model on LiveBench and #1 on Vals Index; far cheaper per task than GPT-6 Astra
  12. Sep 18
    Global KV cache cut to 890 bytes/token, about 437x below DeepSeek V1 and ~4x below V4 Flash
  13. Sep 18
    Scores 74.2 on DeepSWE v1.1 versus GPT-6 Astra's 74, at lower cost and about 4x output speed
  14. Sep 13
    Tied with Kimi K3 and GLM 5.3 on Vals Index, behind on Artificial Analysis, but 3-7x faster via API
  15. Sep 9
    DeepSeek claims V4.1 Flash comprehensively beats V4 Pro across all key metrics
  16. Sep 9
    Found and fixed a flipped comparison operator in a Dockerized air-traffic dashboard without being told, then re-checked the live endpoint eight times
  17. Sep 9
    Reported at 300-400+ tokens per second and 98% of GPT-6 Astra's design benchmark score; unverified
  18. Sep 9
    Unverified third-party numbers put it against Opus 4.8 at max thinking, strong in coding and tool use
  19. Sep 9
    Preview not final: access expected to be removed around September 10 pending post-training; open weights hoped for with the finished version
  20. Aug 21
    The July V4 Flash build (0731) reportedly assumed it could see images and broke sessions trying to read screenshots
  21. Aug 4
    284B-parameter MoE runs fully locally on one Ryzen AI MAX+ 395 with 128GB unified memory, using a 95GB target plus 11GB draft
  22. Aug 4
    In 3D oil-rig test, built the most literal, detailed rig with half-submerged brief met exactly, but more engineering render than landing page
  23. Aug 2
    New 0731 Flash version released, same structure as the DSpark variant; about 10 points above the previous V4 Flash
  24. Aug 2
    Beats MiniMax M3 and the previous V4 Pro, one point below GLM 5.2 on the index shown
  25. Aug 2
    Beats or matches GLM 5.2 and Opus 4.8 on some agentic coding, software engineering and cybersecurity benchmarks
  26. Jul 31
    In one chained-bug full-stack repair task, it fixed all five bugs including a silent Redis cache-key mismatch; reviewer says Fable and the newest GPT failed it

Cost to run

Alternatives

Kimi#1 on Arena webdev among open weights (1684)
GLMCommenters say it feels like Opus 4.8 and handles hard problems better than DeepSeek V4 Flash, though not Fable-class
Qwen MaxFixed all five planted bugs in a Docker/FastAPI/Redis app via Hermes Agent in about 6.5 minutes, with byte-exact self-verification
QwenLocal Qwen3.8-27B Q4 covers leisure coding, but one user still uses DeepSeek 4.1 Flash for real work
More DeepSeek Flash
Built on DeepSeek Flash