Executive readout
The result is quality density. Escha’s 2-bit-class W2 format leads the released 35B HermesAgent field and stays within a few percentage points of the best higher-quant coding runs.
Bottom line. At 12.3 GB of model weights, Escha is 35% smaller than the next-smallest 19.0 GB entry and about one-third the size of a 36.9 GB Q8 model. It posts the best released-model 35B HermesAgent-20 score while landing only 2.4 points behind the best HumanEval+ result, 1.9 behind the best MBPP+ result, and 2.0 behind the best comparable BigCodeBench Hard result.
HermesAgent-20
The headline ranks 12 model-and-quant combinations using each combination’s highest complete 20-scenario score. Bar labels pair quality score with model-weight footprint.
What stands out: Escha’s 90/100 exceeds the next-best released-model result at 88 despite using the lowest-bit weight format in the comparison. Bars are labeled with the quant wherever it is identifiable.
Coding and tool use
HumanEval+ and MBPP+ use EvalPlus plus pass@1, BigCodeBench uses pass@1, and Tool Eval reports its native 100-point score. Each chart keeps only the highest score for a model and quant; bar labels include model-weight footprint.
MBPP scoring is complete. MBPP+: 90.7% base / 75.7% plus. Mbpp/84 remains a preserved failure. The score uses the original 378 first samples.
Released 35B quality ledger
Highest scored result for each public 35B model, quant, and suite, plus the Escha results. Weight footprint is the model-weight artifact size; runtime memory additionally includes context/KV cache and backend workspace.
| Date | Family | Suite | Public release | Quant | Weights | Tasks | Score |
|---|---|---|---|---|---|---|---|
| 20260702T2 | evalplus | humaneval | Qwen3.6-35B-A3B · Chadrock v2 series | ROCmFPX MoEQuality 7.07 BPW | 31.4 GB | 164 | 90.24 |
| 20260702T1 | evalplus | mbpp | Qwen3.6-35B-A3B · Unsloth GGUF | Q6_K_XL | 32.6 GB | 378 | 77.51 |
| 20260702T1 | evalplus | humaneval | Qwen3.6-35B-A3B · Unsloth GGUF | Q6_K_XL | 32.6 GB | 164 | 91.46 |
| 20260626T0 | bigcodebench | bigcodebench-hard-instruct | Ornith 1.0 35B | Q4_K_M | 21.2 GB | 148 | 27.70 |
| 20260607T0 | bigcodebench | bigcodebench-hard-instruct | Qwen3.6-35B-A3B · Chadrock v2 series | ROCmFP4 | 19.0 GB | 148 | 31.76 |
| 20260604T0 | evalplus | humaneval | Qwen3.6-35B-A3B · Chadrock series | Strix Lean mixed quant | 19.0 GB | 164 | 91.46 |
| 20260602T0 | evalplus | mbpp | Qwen3.6-35B-A3B · Chadrock v2 series | ROCmFP4 | 19.0 GB | 378 | 76.98 |
| 20260602T0 | evalplus | humaneval | Qwen3.6-35B-A3B · Chadrock v2 series | ROCmFP4 | 19.0 GB | 164 | 90.85 |
| 20260527T0 | bigcodebench | bigcodebench-hard-instruct | Qwen3.6-35B-A3B · Crown Halo Dynamic | Dynamic mixed quant | 22.6 GB | 148 | 29.05 |
| 20260523T1 | evalplus | humaneval | Qwen3.6-35B-A3B · Crown Halo Dynamic | Dynamic mixed quant | 22.6 GB | 164 | 89.02 |
| 20260523T1 | evalplus | mbpp | Qwen3.6-35B-A3B · Crown Halo Dynamic | Dynamic mixed quant | 22.6 GB | 378 | 74.87 |
| 2026-08-03 | hermesagent-20 | official-20 | Qwen3.6-35B-A3B · Chadrock v2 series | ROCmFP4 | 19.0 GB | 20 | 83.00 |
| 2026-08-03 | hermesagent-20 | official-20 | Escha W2 | Escha W2 / hybrid 2–3b + INT8 | 12.3 GB | 20 | 90.00 |
| 2026-08-03 | evalplus | humaneval | Escha W2 | Escha W2 / hybrid 2–3b + INT8 | 12.3 GB | 164 | 90.85 |
| 2026-08-03 | evalplus | mbpp | Escha W2 | Escha W2 / hybrid 2–3b + INT8 | 12.3 GB | 378 | 75.66 |
| 2026-08-03 | bigcodebench | bigcodebench-hard-instruct | Escha W2 | Escha W2 / hybrid 2–3b + INT8 | 12.3 GB | 148 | 29.73 |
| 2026-08-03 | tool-eval-bench | standard-69 | Escha W2 | Escha W2 / hybrid 2–3b + INT8 | 12.3 GB | 69 | 87.00 |
| 2026-08-03 | tool-eval-bench | hard-15 | Escha W2 | Escha W2 / hybrid 2–3b + INT8 | 12.3 GB | 15 | 80.00 |
| 2026-07-27 | tool-eval-bench | standard-69 | Ornith 1.0 35B | DualView FPX7 + Q8 MTP | 33.5 GB | 69 | 89.13 |
| 2026-07-27 | bigcodebench | bigcodebench-hard-instruct | Ornith 1.0 35B | DualView FPX7 + Q8 MTP | 33.5 GB | 148 | 27.70 |
| 2026-07-27 | evalplus | humaneval | Ornith 1.0 35B | DualView FPX7 + Q8 MTP | 33.5 GB | 164 | 93.29 |
| 2026-07-26 | hermesagent-20 | official-20 | Ornith 1.0 35B | DualView FPX7 + Q8 MTP | 33.5 GB | 20 | 87.00 |
| 2026-07-26 | hermesagent-20 | official-20 | Ornith 1.0 35B | Q7S8 hybrid | 32.6 GB | 20 | 88.00 |
| 2026-07-25 | hermesagent-20 | official-20 | Ornith 1.0 35B | Q8 | 36.9 GB | 20 | 88.00 |
| 2026-07-21 | hermesagent-20 | official-20 | Qwen3.6-35B-A3B | Q8 | 36.9 GB | 20 | 69.00 |
| 2026-07-02 | hermesagent-20 | official-20 | Qwen3.6-35B-A3B · Chadrock v2 series | ROCmFPX MoEQuality 7.07 BPW | 31.4 GB | 20 | 82.00 |
| 2026-07-02 | hermesagent-20 | official-20 | Qwen3.6-35B-A3B · Unsloth GGUF | Q6_K_XL | 32.6 GB | 20 | 71.00 |
| 2026-06-26 | hermesagent-20 | official-20 | Ornith 1.0 35B | Q4_K_M | 21.2 GB | 20 | 82.00 |
| 2026-06-05 | hermesagent-20 | official-20 | Qwen3.6-35B-A3B · Chadrock series | ROCmFP4 | 19.0 GB | 20 | 79.00 |
| 2026-06-05 | hermesagent-20 | official-20 | Qwen3.6-35B-A3B · Crown Halo Dynamic | Dynamic mixed quant | 22.6 GB | 20 | 70.00 |
| 2026-06-05 | hermesagent-20 | official-20 | Qwen3.6-35B-A3B · Chadrock series | Strix Lean mixed quant | 19.0 GB | 20 | 73.00 |
| 2026-05-28 | bfcl | bfcl-v4-all_scoring | Qwen3.6-35B-A3B · Crown Halo Dynamic | Dynamic mixed quant | 22.6 GB | 5217 | 16.86 |
| 2026-05-28 | bfcl | bfcl-v4-non_live | Qwen3.6-35B-A3B · Crown Halo Dynamic | Dynamic mixed quant | 22.6 GB | 1390 | 83.00 |
| 2026-05-28 | livecodebench | release_latest-codegeneration-2025-01-01 | Qwen3.6-35B-A3B · Crown Halo Dynamic | Dynamic mixed quant | 22.6 GB | 182 | 36.81 |