Task ranking

Best Cheap LLM for High Volume

Qwen3.8 27B ranks first for Best Cheap LLM for High Volume as of October 21, 2026.[1][2][3][4] It achieves 3,957.3 tok/s at concurrency 1024 on single H200 GPU with 262K context. Watch out: Disable speculative decoding for prefix-reusing workloads; prefix-cache last-block drop causes 1,648-token recompute and 30-40% throughput loss. DeepSeek V4.1 Flash is the next-ranked alternative. It scores 70.03% on LiveBench Instruction Following (#25 of 59), including paraphrasing, simplifying, story generation, and summarization.

About this recommendation

Updated
Evidence through
Sources
13
Revision
v113

Decision audit

Why this result

Inspect the inputs and the computed order behind the recommendation.

Models screened

24

live candidates

Evaluation feeds

4

task-weighted

Winner coverage

67%

intended feed weight

Largest provider share

2 of 8

deepseek

Established source breadth. 8 citation families and 3 practitioner families provide attributed capability reports; 1 cautionary thread retained. The largest citation family contributes 33%.

Qwen: Qwen3.8 27B leads the measured task comparison with 2 independent evaluation families. Community volume, release age and usage popularity add no ranking points. This category also weights its stated price or context criteria.

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: Qwen3.8 27B
Evaluation feedWeightWinner resultField measured
price weight
67%
83/10024/24
Route reliabilitydisabled; last successful check unknownunavailable
0%
feed unavailable0/24
LiveBench Instruction Following (2026-06-25)Observed 2026-10-18 · evaluation date not supplied · qwen3.8-27bfailing; last successful check 2026-10-18
25%
#1524/24
LMArena TextObserved 2026-10-20 · evaluation date not supplied · Qwen3.8-27B
8%
#6324/24

Provider concentration

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

deepseek25%
  • deepseek2 models
  • Google2 models
  • Z.ai2 models
  • Moonshot AI1 model
  • Qwen1 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. 01Qwen3.8 27B79

    Qwen · coverage 67%

    #15 LiveBench Instruction Following · #63 LMArena Text

  2. 02DeepSeek V4.1 Flash77

    deepseek · coverage 67%

    #25 LiveBench Instruction Following · #26 LMArena Text

  3. 03GLM 5.373

    Z.ai · coverage 67%

    #17 LMArena Text · #27 LiveBench Instruction Following

  4. 04GLM 5.272

    Z.ai · coverage 67%

    #24 LMArena Text · #48 LiveBench Instruction Following

  5. 05Gemini 3.8 Flash69

    Google · coverage 67%

    #1 LiveBench Instruction Following · #8 LMArena Text

  6. 06Kimi K368

    Moonshot AI · coverage 67%

    #14 LMArena Text · #20 LiveBench Instruction Following

  7. 07Gemini 3.7 Flash68

    Google · coverage 67%

    #2 LiveBench Instruction Following · #13 LMArena Text

  8. 08DeepSeek V4 Flash 042366

    deepseek · coverage 67%

    #44 LiveBench Instruction Following · #66 LMArena Text

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

    Qwen3.8 27BQwen · $0.040 in / $1.35 out per M tokens · 1M context · open-weight

    Qwen3.8 27B achieves 3,957.3 tok/s at concurrency 1024 on single H200 GPU with 262K context.

    Best when: Use for maximum throughput on single-GPU deployments; 3,957.3 tok/s at C1024 with 6.88 full-length concurrent requests fits one H200.
    Source 2
    “Qwen3.8-27B is a 27B dense hybrid-attention model with a vision tower and an MTP draft head. It fits one H200 GPU at the full 262144 context, so it queues faster than any recipe that needs a whole node. | | Measured | | --- | --- | | One caller | 122.1 tok/s | | Concurrency 1024 | 3957.3 tok/s, still rising | | Context | 262144, the checkpoint maximum | | KV cache | 1,804,253 tokens, 6.88 full-length requests at once | | Hardware | 1 H200 GPU, TP1, no NCCL | Scope - New recipe at recipes/Qwen3.…”
    mmshadOpen original ↗

    Watch out for

    • Disable speculative decoding for prefix-reusing workloads; prefix-cache last-block drop causes 1,648-token recompute and 30-40% throughput loss.
      Source 3
      “EAGLE/MTP prefix-cache last-block drop causes a 1,648-token recompute per hit on one hybrid Qwen3.8 GDN layout — ~30-40% batch throughput loss on prefix-reusing workloads with speculative decoding enabled Environment - vLLM 0.26.1rc1.dev926+gb05ae5dc0, V1/V2 model runner, async scheduling - Dual RTX 5090 TP=2, FP8 weights + FP8 KV, prefix caching on - Model: Qwen3.8-27B (hybrid: GDN/Mamba layers + full attention → 15 MambaSpec KV groups + full-attention group; scheduler/hash block = 1648 t…”
      Suppressor72Open original ↗
    • Calibrate cost monitoring; char-based token estimation overestimates by ~30% until end-of-message usage calibration.
      Source 5
      “Char-based token estimate deviates ~30% on Qwen-family tokenizers (measured) Summary charsToTokens() assumes 3.0 chars/token for non-CJK and 0.7 tokens/char for CJK. Measured against a Qwen3.8 tokenizer (Qwen3.8-27B served locally, usage counts exact output tokens including reasoning), both constants are off by ~25–35% in the same direction (overestimate), so mid-stream inst/avg read ~25–35% above the true rate until the end-of-message usage calibration kicks in. Measu…”
      bioShaunOpen original ↗
  2. 02

    DeepSeek V4.1 Flashdeepseek · $0.049 in / $0.38 out per M tokens · 1.0M context · open-weight

    DeepSeek V4.1 Flash scores 70.03% on LiveBench Instruction Following (#25 of 59), including paraphrasing, simplifying, story generation, and summarization.

    Best when: Its strongest result for this task is on LiveBench Instruction Following, which measures following detailed instructions exactly.
    Source 4
    “Within the 59 catalog models matched to this evaluation: Scores 70.03% on LiveBench Instruction Following (#25 of 59), including paraphrasing, simplifying, story generation, and summarization.”
    LiveBench Instruction FollowingOpen original ↗
  3. 03

    GLM 5.3Z.ai · $0.079 in / $1.40 out per M tokens · 1.0M context · open-weight

    GLM 5.3 ranks #17 of 150 on LMArena's overall text arena (Elo 1481), based on blind human preference votes.

    Best when: Its strongest result for this task is on LMArena Text, which measures general reasoning and instruction following.
    Source 6
    “Within the 150 catalog models matched to this evaluation: Ranks #17 of 150 on LMArena's overall text arena (Elo 1481), based on blind human preference votes.”
    LMArena text arenaOpen original ↗

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

  4. 04

    GLM 5.2Z.ai · $0.034 in / $7 out per M tokens · 1.0M context · open-weight

    GLM 5.2 ranks #24 of 150 on LMArena's overall text arena (Elo 1476), based on blind human preference votes.

    Best when: Its strongest result for this task is on LMArena Text, which measures general reasoning and instruction following.
    Source 7
    “Within the 150 catalog models matched to this evaluation: Ranks #24 of 150 on LMArena's overall text arena (Elo 1476), based on blind human preference votes.”
    LMArena text arenaOpen original ↗

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

  5. 05

    Gemini 3.8 FlashGoogle · $0.83 in / $4.13 out per M tokens · 1.0M context

    Gemini 3.8 Flash scores 81.41% on LiveBench Instruction Following (#1 of 59), including paraphrasing, simplifying, story generation, and summarization.

    Best when: Its strongest result for this task is on LiveBench Instruction Following, which measures following detailed instructions exactly.
    Source 8
    “Within the 59 catalog models matched to this evaluation: Scores 81.41% on LiveBench Instruction Following (#1 of 59), including paraphrasing, simplifying, story generation, and summarization.”
    LiveBench Instruction FollowingOpen original ↗
  6. 06

    Kimi K3Moonshot AI · $0.24 in / $11.94 out per M tokens · 1.0M context · open-weight

    Kimi K3 ranks #14 of 150 on LMArena's overall text arena (Elo 1488), based on blind human preference votes.

    Best when: Its strongest result for this task is on LMArena Text, which measures general reasoning and instruction following.
    Source 9
    “Within the 150 catalog models matched to this evaluation: Ranks #14 of 150 on LMArena's overall text arena (Elo 1488), based on blind human preference votes.”
    LMArena text arenaOpen original ↗

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

  7. 07

    Gemini 3.7 FlashGoogle · $0.75 in / $3.75 out per M tokens · 1.0M context

    Gemini 3.7 Flash scores 79.93% on LiveBench Instruction Following (#2 of 59), including paraphrasing, simplifying, story generation, and summarization.

    Best when: Its strongest result for this task is on LiveBench Instruction Following, which measures following detailed instructions exactly.
    Source 10
    “Within the 59 catalog models matched to this evaluation: Scores 79.93% on LiveBench Instruction Following (#2 of 59), including paraphrasing, simplifying, story generation, and summarization.”
    LiveBench Instruction FollowingOpen original ↗

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

  8. 08

    DeepSeek V4 Flash 0423deepseek · $0.091 in / $0.18 out per M tokens · 1.0M context · open-weight

    DeepSeek V4 Flash 0423 ranks #66 on LMArena (Elo 1436) with peak pricing at $0.014/$0.44/$1.32 per 1M tokens.

    Best when: Its strongest result for this task is on LiveBench Instruction Following, which measures following detailed instructions exactly.
    Source 11
    “Within the 59 catalog models matched to this evaluation: Scores 63.14% on LiveBench Instruction Following (#44 of 59), including paraphrasing, simplifying, story generation, and summarization.”
    LiveBench Instruction FollowingOpen original ↗

    Watch out for

    • Watch for catastrophic cache failures; quota exhausted in ~20 minutes after cache reads suddenly dropped to 0.
      Source 12
      “This issue might be a duplicate of existing issues. Please check: - #41125: OpenCode Go endpoint: no prompt caching on DeepSeek (same symptom — cache_hit always 0 on repeated identical prompts via the Go endpoint; you already referenced this one) - #42935: OpenCode Go quota exhausted in ~20 minutes after DeepSeek V4 Flash cache reads suddenly dropped to 0 (same root cause, quota impact) - #43218: Opencodego cache hit rate extremely low (<10%) when used via ccSwitch with Claude Code / Codex (bro…”
      github-actions[bot]Open original ↗
    • Verify MI35x ROCm compatibility; failed accuracy and performance tests with FP8+FP4 quantization.
      Source 13
      “nightly-test-amd.yml — Per-Job Analysis (continued 9/12) nightly-8-gpu-mi35x-deepseek-v4-flash (rocm10, linux-mi35x-gpu-8) — failed step(s): Accuracy + Performance Test MI35x ROCm (8-GPU DeepSeek-V4-Flash FP8 + FP4, unified_kv_triton) Commit Info - sglang: b6778515 — "[Fix] glm5.2: Align HiSparse warmup backends and move test to nightly ([#44303](https://github.com/sgl-project/sglang/pull/443…”
      bingxcheOpen original ↗

Frequently asked

What is the best model for Best Cheap LLM for High Volume?
Qwen3.8 27B ranks first as of October 21, 2026. Qwen3.8 27B achieves 3,957.3 tok/s at concurrency 1024 on single H200 GPU with 262K context.[1][2]
What should I watch out for with Qwen3.8 27B?
Disable speculative decoding for prefix-reusing workloads; prefix-cache last-block drop causes 1,648-token recompute and 30-40% throughput loss.[3]
What is the best alternative to Qwen3.8 27B?
DeepSeek V4.1 Flash ranks second. DeepSeek V4.1 Flash scores 70.03% on LiveBench Instruction Following (#25 of 59), including paraphrasing, simplifying, story generation, and summarization.[4]
Can I self-host the top pick for Best Cheap LLM for High Volume?
Yes. Qwen3.8 27B is open-weight, so it can run on your own hardware as well as through hosted routes.[1][2]

Sources

  1. 1

    “Within the 150 catalog models matched to this evaluation: Ranks #63 of 150 on LMArena's overall text arena (Elo 1437), based on blind human preference votes.”

    LMArena text arena · Benchmark · Oct 20, 2026
  2. 2

    “Qwen3.8-27B is a 27B dense hybrid-attention model with a vision tower and an MTP draft head. It fits one H200 GPU at the full 262144 context, so it queues faster than any recipe that needs a whole node. | | Measured | | --- | --- | | One caller | 122.1 tok/s | | Concurrency 1024 | 3957.3 tok/s, still rising | | Context | 262144, the checkpoint maximum | | KV cache | 1,804,253 tokens, 6.88 full-length requests at once | | Hardware | 1 H200 GPU, TP1, no NCCL | Scope - New recipe at recipes/Qwen3.…”

    mmshad · GitHub · Aug 26, 2026
  3. 3

    “EAGLE/MTP prefix-cache last-block drop causes a 1,648-token recompute per hit on one hybrid Qwen3.8 GDN layout — ~30-40% batch throughput loss on prefix-reusing workloads with speculative decoding enabled Environment - vLLM 0.26.1rc1.dev926+gb05ae5dc0, V1/V2 model runner, async scheduling - Dual RTX 5090 TP=2, FP8 weights + FP8 KV, prefix caching on - Model: Qwen3.8-27B (hybrid: GDN/Mamba layers + full attention → 15 MambaSpec KV groups + full-attention group; scheduler/hash block = 1648 t…”

    Suppressor72 · GitHub · Aug 25, 2026
  4. 4

    “Within the 59 catalog models matched to this evaluation: Scores 70.03% on LiveBench Instruction Following (#25 of 59), including paraphrasing, simplifying, story generation, and summarization.”

    LiveBench Instruction Following · Benchmark · Oct 18, 2026
  5. 5

    “Char-based token estimate deviates ~30% on Qwen-family tokenizers (measured) Summary charsToTokens() assumes 3.0 chars/token for non-CJK and 0.7 tokens/char for CJK. Measured against a Qwen3.8 tokenizer (Qwen3.8-27B served locally, usage counts exact output tokens including reasoning), both constants are off by ~25–35% in the same direction (overestimate), so mid-stream inst/avg read ~25–35% above the true rate until the end-of-message usage calibration kicks in. Measu…”

    bioShaun · GitHub · Aug 24, 2026
  6. 6

    “Within the 150 catalog models matched to this evaluation: Ranks #17 of 150 on LMArena's overall text arena (Elo 1481), based on blind human preference votes.”

    LMArena text arena · Benchmark · Oct 20, 2026
  7. 7

    “Within the 150 catalog models matched to this evaluation: Ranks #24 of 150 on LMArena's overall text arena (Elo 1476), based on blind human preference votes.”

    LMArena text arena · Benchmark · Oct 20, 2026
  8. 8

    “Within the 59 catalog models matched to this evaluation: Scores 81.41% on LiveBench Instruction Following (#1 of 59), including paraphrasing, simplifying, story generation, and summarization.”

    LiveBench Instruction Following · Benchmark · Oct 18, 2026
  9. 9

    “Within the 150 catalog models matched to this evaluation: Ranks #14 of 150 on LMArena's overall text arena (Elo 1488), based on blind human preference votes.”

    LMArena text arena · Benchmark · Oct 20, 2026
  10. 10

    “Within the 59 catalog models matched to this evaluation: Scores 79.93% on LiveBench Instruction Following (#2 of 59), including paraphrasing, simplifying, story generation, and summarization.”

    LiveBench Instruction Following · Benchmark · Oct 18, 2026
  11. 11

    “Within the 59 catalog models matched to this evaluation: Scores 63.14% on LiveBench Instruction Following (#44 of 59), including paraphrasing, simplifying, story generation, and summarization.”

    LiveBench Instruction Following · Benchmark · Oct 18, 2026
  12. 12

    “This issue might be a duplicate of existing issues. Please check: - #41125: OpenCode Go endpoint: no prompt caching on DeepSeek (same symptom — cache_hit always 0 on repeated identical prompts via the Go endpoint; you already referenced this one) - #42935: OpenCode Go quota exhausted in ~20 minutes after DeepSeek V4 Flash cache reads suddenly dropped to 0 (same root cause, quota impact) - #43218: Opencodego cache hit rate extremely low (<10%) when used via ccSwitch with Claude Code / Codex (bro…”

    github-actions[bot] · GitHub · Sep 9, 2026
  13. 13

    “nightly-test-amd.yml — Per-Job Analysis (continued 9/12) nightly-8-gpu-mi35x-deepseek-v4-flash (rocm10, linux-mi35x-gpu-8) — failed step(s): Accuracy + Performance Test MI35x ROCm (8-GPU DeepSeek-V4-Flash FP8 + FP4, unified_kv_triton) Commit Info - sglang: b6778515 — "[Fix] glm5.2: Align HiSparse warmup backends and move test to nightly ([#44303](https://github.com/sgl-project/sglang/pull/443…”

    bingxche · GitHub · Oct 15, 2026

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