Part of Best LLM for Coding

Best Local LLM for Coding

DeepSeek V4.1 Flash 1017 is the provisional first pick for Best Local LLM for Coding as of November 15, 2026, while independent task measurements are still limited.[1][2][3][4][5][6][7] It ranks #15 of 102 on LMArena's WebDev coding arena (Elo 1664), a leaderboard built from blind human preference votes on coding tasks. GLM 5.2 is the next-ranked alternative. It resolves 62.88% on SWE-rebench (#4 of 12), ranks #30 of 102 on LMArena's WebDev coding arena (Elo 1608), scores 79.65% on LiveBench Coding (#19 of 59).

About this recommendation

Updated
Evidence through
Sources
17
Revision
v144

Decision audit

Why this result

Inspect the inputs and the computed order behind the recommendation.

Models screened

24

live candidates

Evaluation feeds

5

task-weighted

Winner coverage

29%

intended feed weight

Largest provider share

3 of 8

Z.ai

Provisional source breadth. 12 citation families and 0 practitioner families provide attributed capability reports; 0 cautionary threads retained. The largest citation family contributes 36%.

DeepSeek: DeepSeek V4.1 Flash 1017 leads the measured task comparison with limited independent task measurements. Community volume, release age and usage popularity add no ranking points. This category also weights its stated price or context criteria.

Real-world track record is still developing. The recommendation can lead on independent measurements while practitioner reports accumulate.

Sources evaluated

These are the weights used in scoring. Missing measurements are omitted from each model’s average. A feed with no result for any model released in the last 60 days is stale and does not score until it catches up.

winner: DeepSeek V4.1 Flash 1017
Evaluation feedWeightWinner resultField measured
Aider Polyglotstale
0%
no current result0/24
LiveBench Codingfailing; last successful check 2026-10-18stale
0%
no current result0/24
LMArena WebDevObserved 2026-11-11 · evaluation date not supplied · max effort
64%
#1524/24
SWE-rebenchstale
0%
no current result0/24
price weight
36%
90/10024/24

Provider concentration

Each exact model is scored separately; provider identity is not a ranking input.

Z.ai38%
  • Z.ai3 models
  • deepseek2 models
  • Moonshot AI2 models
  • xiaomi1 model

Decision table

Every published model is shown in computed order. Practitioner sources are distinct community threads, not the citations repeated in the prose below.

  1. deepseek · coverage 29% · provisional evidence

    #15 LMArena WebDev

  2. Z.ai · coverage 29% · provisional evidence

    #30 LMArena WebDev

  3. deepseek · coverage 29% · provisional evidence

    #26 LMArena WebDev

  4. Z.ai · coverage 29% · provisional evidence

    #28 LMArena WebDev

  5. Moonshot AI · coverage 29% · provisional evidence

    #10 LMArena WebDev

  6. Z.ai · coverage 29% · provisional evidence

    #25 LMArena WebDev

  7. Moonshot AI · coverage 29% · provisional evidence

    #16 LMArena WebDev

  8. xiaomi · coverage 29% · provisional evidence

    #21 LMArena WebDev

Task score averages the available, task-weighted measurements. Benchmark positions compare models matched to our catalog. Models measured by at least two independent evaluator families are considered first. Missing tests, practitioner report counts, popularity and release age add no points. Coverage describes evidence breadth, not a probability of correctness. Scores within two points are shown as close under this policy; this is not a statistical confidence interval. When the leading scores are that close, the model measured by more evaluator families ranks first.

  1. 01

    DeepSeek V4.1 Flash 1017deepseek · $0.020 in / $0.11 out per M tokens · 1.0M context · open-weight

    DeepSeek V4.1 Flash 1017 ranks #15 of 102 on LMArena's WebDev coding arena (Elo 1664), a leaderboard built from blind human preference votes on coding tasks.

    Best when: Its strongest result for this task is on LMArena WebDev, which measures practical, real-world coding quality.
    Source 1
    “Within the 102 catalog models matched to this evaluation: Ranks #15 of 102 on LMArena's WebDev coding arena (Elo 1664), a leaderboard built from blind human preference votes on coding tasks.”
    LMArena WebDev (coding) arenaOpen original ↗

    No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.

  2. 02

    GLM 5.2Z.ai · $0.007 in / $6.80 out per M tokens · 1.0M context · open-weight

    GLM 5.2 resolves 62.88% on SWE-rebench (#4 of 12), ranks #30 of 102 on LMArena's WebDev coding arena (Elo 1608), scores 79.65% on LiveBench Coding (#19 of 59).

    Best when: Its strongest result for this task is on LMArena WebDev, which measures practical, real-world coding quality.
    Source 3
    “Within the 102 catalog models matched to this evaluation: Ranks #30 of 102 on LMArena's WebDev coding arena (Elo 1608), a leaderboard built from blind human preference votes on coding tasks.”
    LMArena WebDev (coding) arenaOpen original ↗

    Watch out for

    • Expect impractical inference speeds around 0.15 tok/s on stock builds or ~1 tok/s on Nvidia Spark hardware, far below usable interactive coding thresholds.
      Source 5
      “Hi @VincentMarquez , Thanks for the great reading, I tried to experiment running GLM-5.2 on an nvidia Spark using colibri and all your recommendations, but I can't reach anything remotely close to your results. I'm actually reaching anything close to 1 tok/s ... I took the advised pointed to me by @mohamedmastouri2000-boop here ( https://github.com/JustVugg/colibri/issues/161 ) but it's not helping... What am I doing wrong there?”
      Source 6
      “Commit 81f08a09e5651ce52616dc720f68810f9021c0be Hardware and storage Hi there, I use Opus 4.8 to drive the benchmark, it add the following benchmark results to the docs/benchmarks.md, appreciate if you have any idea to improve the performance: Community benchmarks (measured) Real numbers from real machines, stock build (setup.sh, gcc 13), greedy decoding, --ngen 32, MTP active: | machine | disk (iobench, 19 MB × 64, 8 threads) | config | measured | |---|---|---|---| | Intel i5-…”
    • Accept substantial accuracy degradation if running FP2 quantization (the only viable home configuration), with over 17% loss making it perform worse than Claude Sonnet 4.6.
      Source 7
      “"GLM 5.2 is just shy of GPT 5.4"... If your running the full model. As in have 750 (FP8) to 1.5TB(FP16) of memory available. Do not mix the benchmark results of GLM 5.2 FP16 FP8 with FP4 or FP2. * FP4 will mean a accuracy loss of about 3%. Not noticeable but more chance for mistakes. * FP2 ... what is what most people are able to run at home, for a "reasonable" price. Your looking at over 17% loss in accuracy. At that point, your running at less then claude-sonnet-4.6, as the issues compound wi…”
  3. 03

    DeepSeek V4.1 Flashdeepseek · $0.051 in / $0.20 out per M tokens · 1.0M context · open-weight

    DeepSeek V4.1 Flash ranks #26 of 102 on LMArena's WebDev coding arena (Elo 1621), scores 80.04% on LiveBench Coding (#18 of 59), achieves 55.2 tok/s decode with DSpark on MLX engine.

    Best when: Deploy with MLX engine and ISTA's GGUF sets for fast decoding at 55.2 tok/s on Apple Silicon, outperforming llama.cpp-based engines.
    Source 8
    “Before opening - [x] I searched open and closed issues, open PRs and recent commits for this, and it is not already reported or fixed Problem --mlx-gguf serves ISTA's DeepSeek-V4.1-Flash GSQ-RCO sets on the MLX engine, and it is fast: 55.2 tok/s decode with DSpark on the IQ2_XS set (M3 Ultra, 512 GB, mlx-serve 26.11.2), where llama.cpp-based engines give 33 to 37 on the same files. The same flag declines unsloth/DeepSeek-V4.1-Flash-1017-GGUF UD-Q2_K_XL. The debug log says: The set t…”

    More tips

    • Target for code generation tasks where LiveBench scores above 80% and LMArena rankings place it in the top quarter of models.
      Source 9
      “Within the 102 catalog models matched to this evaluation: Ranks #26 of 102 on LMArena's WebDev coding arena (Elo 1621), a leaderboard built from blind human preference votes on coding tasks.”
      LMArena WebDev (coding) arenaOpen original ↗
      Source 10
      “Within the 59 catalog models matched to this evaluation: Scores 80.04% on LiveBench Coding (#18 of 59), an objective evaluation of code generation and completion.”
      LiveBench CodingOpen original ↗
  4. 04

    GLM 5.3 FlashZ.ai · $0.040 in / $0.078 out per M tokens · 1.0M context · open-weight

    GLM 5.3 Flash ranks #28 of 102 on LMArena's WebDev coding arena (Elo 1609), scores 78.95% on LiveBench Coding (#22 of 59).

    Best when: Benchmark for web development and code completion tasks where its LMArena and LiveBench standings suggest adequate capability.
    Source 11
    “Within the 102 catalog models matched to this evaluation: Ranks #28 of 102 on LMArena's WebDev coding arena (Elo 1609), a leaderboard built from blind human preference votes on coding tasks.”
    LMArena WebDev (coding) arenaOpen original ↗
    Source 12
    “Within the 59 catalog models matched to this evaluation: Scores 78.95% on LiveBench Coding (#22 of 59), an objective evaluation of code generation and completion.”
    LiveBench CodingOpen original ↗

    No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.

  5. 05

    Kimi K3.1 ProMoonshot AI · $0.24 in / $8.40 out per M tokens · 1.0M context · open-weight

    Kimi K3.1 Pro ranks #10 of 102 on LMArena's WebDev coding arena (Elo 1687), a leaderboard built from blind human preference votes on coding tasks.

    Best when: Its strongest result for this task is on LMArena WebDev, which measures practical, real-world coding quality.
    Source 13
    “Within the 102 catalog models matched to this evaluation: Ranks #10 of 102 on LMArena's WebDev coding arena (Elo 1687), a leaderboard built from blind human preference votes on coding tasks.”
    LMArena WebDev (coding) arenaOpen original ↗

    No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.

  6. 06

    GLM 5.3Z.ai · $0.084 in / $3.60 out per M tokens · 1.0M context · open-weight

    GLM 5.3 ranks #25 of 102 on LMArena's WebDev coding arena (Elo 1622) and scores 78.95% on LiveBench Coding (#22 of 59).

    Best when: Use for code generation and completion where LiveBench scores near 79% and LMArena rankings indicate solid web development performance.
    Source 14
    “Within the 102 catalog models matched to this evaluation: Ranks #25 of 102 on LMArena's WebDev coding arena (Elo 1622), a leaderboard built from blind human preference votes on coding tasks.”
    LMArena WebDev (coding) arenaOpen original ↗
    Source 15
    “Within the 59 catalog models matched to this evaluation: Scores 78.95% on LiveBench Coding (#22 of 59), an objective evaluation of code generation and completion.”
    LiveBench CodingOpen original ↗

    No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.

  7. 07

    Kimi K3Moonshot AI · $0.19 in / $16 out per M tokens · 1.0M context · open-weight

    Kimi K3 ranks #16 of 102 on LMArena's WebDev coding arena (Elo 1658), a leaderboard built from blind human preference votes on coding tasks.

    Best when: Its strongest result for this task is on LMArena WebDev, which measures practical, real-world coding quality.
    Source 16
    “Within the 102 catalog models matched to this evaluation: Ranks #16 of 102 on LMArena's WebDev coding arena (Elo 1658), a leaderboard built from blind human preference votes on coding tasks.”
    LMArena WebDev (coding) arenaOpen original ↗

    No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.

  8. 08

    MiMo-V2.6-Flashxiaomi · $0.14 in / $0.28 out per M tokens · 1.1M context · open-weight

    MiMo-V2.6-Flash ranks #21 of 102 on LMArena's WebDev coding arena with an Elo of 1638, offering a mid-tier placement for local coding use.

    Best when: Evaluate for web development tasks where LMArena human preference data indicates acceptable performance relative to 81 other catalog models.
    Source 17
    “Within the 102 catalog models matched to this evaluation: Ranks #21 of 102 on LMArena's WebDev coding arena (Elo 1638), a leaderboard built from blind human preference votes on coding tasks.”
    LMArena WebDev (coding) arenaOpen original ↗

    No practitioner reports are linked yet, so no cautions are listed. Absence of reports is not evidence of reliability.

Frequently asked

What is the best model for Best Local LLM for Coding?
DeepSeek V4.1 Flash 1017 ranks first as of November 15, 2026. DeepSeek V4.1 Flash 1017 ranks #15 of 102 on LMArena's WebDev coding arena (Elo 1664), a leaderboard built from blind human preference votes on coding tasks.[1]
What is the best alternative to DeepSeek V4.1 Flash 1017?
GLM 5.2 ranks second. GLM 5.2 resolves 62.88% on SWE-rebench (#4 of 12), ranks #30 of 102 on LMArena's WebDev coding arena (Elo 1608), scores 79.65% on LiveBench Coding (#19 of 59).[2][3][4][5][6][7]
Can I self-host the top pick for Best Local LLM for Coding?
Yes. DeepSeek V4.1 Flash 1017 is open-weight, so it can run on your own hardware as well as through hosted routes.[1]

Sources

  1. 1

    “Within the 102 catalog models matched to this evaluation: Ranks #15 of 102 on LMArena's WebDev coding arena (Elo 1664), a leaderboard built from blind human preference votes on coding tasks.”

    LMArena WebDev (coding) arena · Benchmark · Nov 11, 2026
  2. 2

    “Within the 12 catalog models matched to this evaluation: Resolves 62.88% ± 1.19 standard error on SWE-rebench (#4 of 12) using tools, comparing the same task window (2026-05-15 to 2026-07-01) and agent setup across models.”

    SWE-rebench · Benchmark · Nov 15, 2026
  3. 3

    “Within the 102 catalog models matched to this evaluation: Ranks #30 of 102 on LMArena's WebDev coding arena (Elo 1608), a leaderboard built from blind human preference votes on coding tasks.”

    LMArena WebDev (coding) arena · Benchmark · Nov 11, 2026
  4. 4

    “Within the 59 catalog models matched to this evaluation: Scores 79.65% on LiveBench Coding (#19 of 59), an objective evaluation of code generation and completion.”

    LiveBench Coding · Benchmark · Oct 18, 2026
  5. 5

    “Hi @VincentMarquez , Thanks for the great reading, I tried to experiment running GLM-5.2 on an nvidia Spark using colibri and all your recommendations, but I can't reach anything remotely close to your results. I'm actually reaching anything close to 1 tok/s ... I took the advised pointed to me by @mohamedmastouri2000-boop here ( https://github.com/JustVugg/colibri/issues/161 ) but it's not helping... What am I doing wrong there?”

    RoiArthurB · GitHub · Aug 1, 2026
  6. 6

    “Commit 81f08a09e5651ce52616dc720f68810f9021c0be Hardware and storage Hi there, I use Opus 4.8 to drive the benchmark, it add the following benchmark results to the docs/benchmarks.md, appreciate if you have any idea to improve the performance: Community benchmarks (measured) Real numbers from real machines, stock build (setup.sh, gcc 13), greedy decoding, --ngen 32, MTP active: | machine | disk (iobench, 19 MB × 64, 8 threads) | config | measured | |---|---|---|---| | Intel i5-…”

    limcheekin · GitHub · Jul 24, 2026
  7. 7

    “"GLM 5.2 is just shy of GPT 5.4"... If your running the full model. As in have 750 (FP8) to 1.5TB(FP16) of memory available. Do not mix the benchmark results of GLM 5.2 FP16 FP8 with FP4 or FP2. * FP4 will mean a accuracy loss of about 3%. Not noticeable but more chance for mistakes. * FP2 ... what is what most people are able to run at home, for a "reasonable" price. Your looking at over 17% loss in accuracy. At that point, your running at less then claude-sonnet-4.6, as the issues compound wi…”

    benjiro29 · Hacker News · Jun 22, 2026
  8. 8

    “Before opening - [x] I searched open and closed issues, open PRs and recent commits for this, and it is not already reported or fixed Problem --mlx-gguf serves ISTA's DeepSeek-V4.1-Flash GSQ-RCO sets on the MLX engine, and it is fast: 55.2 tok/s decode with DSpark on the IQ2_XS set (M3 Ultra, 512 GB, mlx-serve 26.11.2), where llama.cpp-based engines give 33 to 37 on the same files. The same flag declines unsloth/DeepSeek-V4.1-Flash-1017-GGUF UD-Q2_K_XL. The debug log says: The set t…”

    soulmachine · GitHub · Nov 12, 2026
  9. 9

    “Within the 102 catalog models matched to this evaluation: Ranks #26 of 102 on LMArena's WebDev coding arena (Elo 1621), a leaderboard built from blind human preference votes on coding tasks.”

    LMArena WebDev (coding) arena · Benchmark · Nov 11, 2026
  10. 10

    “Within the 59 catalog models matched to this evaluation: Scores 80.04% on LiveBench Coding (#18 of 59), an objective evaluation of code generation and completion.”

    LiveBench Coding · Benchmark · Oct 18, 2026
  11. 11

    “Within the 102 catalog models matched to this evaluation: Ranks #28 of 102 on LMArena's WebDev coding arena (Elo 1609), a leaderboard built from blind human preference votes on coding tasks.”

    LMArena WebDev (coding) arena · Benchmark · Nov 11, 2026
  12. 12

    “Within the 59 catalog models matched to this evaluation: Scores 78.95% on LiveBench Coding (#22 of 59), an objective evaluation of code generation and completion.”

    LiveBench Coding · Benchmark · Oct 18, 2026
  13. 13

    “Within the 102 catalog models matched to this evaluation: Ranks #10 of 102 on LMArena's WebDev coding arena (Elo 1687), a leaderboard built from blind human preference votes on coding tasks.”

    LMArena WebDev (coding) arena · Benchmark · Nov 11, 2026
  14. 14

    “Within the 102 catalog models matched to this evaluation: Ranks #25 of 102 on LMArena's WebDev coding arena (Elo 1622), a leaderboard built from blind human preference votes on coding tasks.”

    LMArena WebDev (coding) arena · Benchmark · Nov 11, 2026
  15. 15

    “Within the 59 catalog models matched to this evaluation: Scores 78.95% on LiveBench Coding (#22 of 59), an objective evaluation of code generation and completion.”

    LiveBench Coding · Benchmark · Oct 18, 2026
  16. 16

    “Within the 102 catalog models matched to this evaluation: Ranks #16 of 102 on LMArena's WebDev coding arena (Elo 1658), a leaderboard built from blind human preference votes on coding tasks.”

    LMArena WebDev (coding) arena · Benchmark · Nov 11, 2026
  17. 17

    “Within the 102 catalog models matched to this evaluation: Ranks #21 of 102 on LMArena's WebDev coding arena (Elo 1638), a leaderboard built from blind human preference votes on coding tasks.”

    LMArena WebDev (coding) arena · Benchmark · Nov 11, 2026

Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.