"With 65.2% accuracy on HumanEval for code generation, DeepSeek V3 sets a new standard for AI performance."
at google, the job title for a software engineer who writes 35% incorrect code is "former software engineer"
at google, the job title for a software engineer who writes 35% incorrect code is "former software engineer"
March 25, 2025 at 10:46 PM
"With 65.2% accuracy on HumanEval for code generation, DeepSeek V3 sets a new standard for AI performance."
at google, the job title for a software engineer who writes 35% incorrect code is "former software engineer"
at google, the job title for a software engineer who writes 35% incorrect code is "former software engineer"
Salesforce released the open-weight of CoDA-1.7B: a text diffusion coding model that outputs tokens bidirectionally in parallel.
⚡️ Faster inference, 1.7B rivaling 7B.
📊 54.3% HumanEval | 47.6% HumanEval+ | 55.4% EvalPlus
Model: huggingface.co/Salesforce/C...
Report: github.com/SalesforceAI...
⚡️ Faster inference, 1.7B rivaling 7B.
📊 54.3% HumanEval | 47.6% HumanEval+ | 55.4% EvalPlus
Model: huggingface.co/Salesforce/C...
Report: github.com/SalesforceAI...
Salesforce/CoDA-v0-Instruct · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
October 5, 2025 at 2:18 PM
Salesforce released the open-weight of CoDA-1.7B: a text diffusion coding model that outputs tokens bidirectionally in parallel.
⚡️ Faster inference, 1.7B rivaling 7B.
📊 54.3% HumanEval | 47.6% HumanEval+ | 55.4% EvalPlus
Model: huggingface.co/Salesforce/C...
Report: github.com/SalesforceAI...
⚡️ Faster inference, 1.7B rivaling 7B.
📊 54.3% HumanEval | 47.6% HumanEval+ | 55.4% EvalPlus
Model: huggingface.co/Salesforce/C...
Report: github.com/SalesforceAI...
The quality of Qwen2.5-Code-32B-Instruct remains virtually unaffected by quantization.
"HumanEval benchmark of EXL2 quants of popular local LLMs (2.5 through 8.0 bpw covered)"
Link: www.reddit.com/r/LocalLLaMA...
"HumanEval benchmark of EXL2 quants of popular local LLMs (2.5 through 8.0 bpw covered)"
Link: www.reddit.com/r/LocalLLaMA...
November 17, 2024 at 11:16 PM
The quality of Qwen2.5-Code-32B-Instruct remains virtually unaffected by quantization.
"HumanEval benchmark of EXL2 quants of popular local LLMs (2.5 through 8.0 bpw covered)"
Link: www.reddit.com/r/LocalLLaMA...
"HumanEval benchmark of EXL2 quants of popular local LLMs (2.5 through 8.0 bpw covered)"
Link: www.reddit.com/r/LocalLLaMA...
Meanwhile, further independent evaluation confirms SYNTH performs better than Nanochat/Fineweb across nearly all benchmarks.
(Humaneval to be expected: we did not out any code inside yet)
(Humaneval to be expected: we did not out any code inside yet)
Outside of HumanEval, the model pretrained on the SYNTH dataset outperforms the standard nanochat on every task. I report results for both the standard chat template (Harmony) as well as the Qwen3 Chat template used in the SYNTH dataset.
February 1, 2026 at 10:50 PM
Meanwhile, further independent evaluation confirms SYNTH performs better than Nanochat/Fineweb across nearly all benchmarks.
(Humaneval to be expected: we did not out any code inside yet)
(Humaneval to be expected: we did not out any code inside yet)
Announcing CheatGPT, a revolutionary model that achieves SoTA on HumanEval! It's incredibly sample-efficient – just ONE training sample – and *tiny*, fitting on your Casio wristwatch!
![Snippet of python code implementing HumanEval's generate_one_completion(prompt) harness, and output from the HumanEval benchmark runner showing a pass@1 score of 94.5%.
[spoiler] The trick is that the "model" returns a Python object that overrides the equality operator and always returns true, causing the tests in the HumanEval test suite to pass.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:r5q2jrgxqoa2ovw6lz667it7/bafkreiem52ra4fbn5qc33rehniup2tcz26pqkbzg4kc76drpnj4i64ev3u)
April 1, 2025 at 7:19 PM
Announcing CheatGPT, a revolutionary model that achieves SoTA on HumanEval! It's incredibly sample-efficient – just ONE training sample – and *tiny*, fitting on your Casio wristwatch!
Alibaba's Qwen2.5-Coder-32B-Instruct
The flagship model, Qwen2.5-Coder-32B-Instruct, reaches 92.7 in HumanEval, 90.2 in MBPP, 31.4 in LiveCodeBench, 73.7 in Aider, 85.1 in Spider, and 68.9 in CodeArena!
The flagship model, Qwen2.5-Coder-32B-Instruct, reaches 92.7 in HumanEval, 90.2 in MBPP, 31.4 in LiveCodeBench, 73.7 in Aider, 85.1 in Spider, and 68.9 in CodeArena!
November 11, 2024 at 6:31 PM
Alibaba's Qwen2.5-Coder-32B-Instruct
The flagship model, Qwen2.5-Coder-32B-Instruct, reaches 92.7 in HumanEval, 90.2 in MBPP, 31.4 in LiveCodeBench, 73.7 in Aider, 85.1 in Spider, and 68.9 in CodeArena!
The flagship model, Qwen2.5-Coder-32B-Instruct, reaches 92.7 in HumanEval, 90.2 in MBPP, 31.4 in LiveCodeBench, 73.7 in Aider, 85.1 in Spider, and 68.9 in CodeArena!
Verified hand-written Lean solutions to HumanEval. ~ Markus Himmel. github.com/TwoFX/human-... #ITP #LeanProver
GitHub - TwoFX/human-eval-lean: Hand-written verified Lean solutions for the HumanEval benchmark
Hand-written verified Lean solutions for the HumanEval benchmark - TwoFX/human-eval-lean
github.com
April 30, 2025 at 11:11 AM
Verified hand-written Lean solutions to HumanEval. ~ Markus Himmel. github.com/TwoFX/human-... #ITP #LeanProver
On the the most commonly-used benchmarks is MMLU, which multiple choice about knowledge across lots of fields. Others include writing programs in Python (HumanEval) or solving reasoning puzzles involving recognizing patterns (ARC-AGI), or even taking standardized tests like the LSAT.
January 27, 2025 at 4:53 PM
On the the most commonly-used benchmarks is MMLU, which multiple choice about knowledge across lots of fields. Others include writing programs in Python (HumanEval) or solving reasoning puzzles involving recognizing patterns (ARC-AGI), or even taking standardized tests like the LSAT.
Looking for evidence of data leakage in the #HumanEval code generation #LLM benchmark? Check out our analysis of a subset of HumanEval tasks vs comparable tasks on #ChatGPT, #Claude and #Llama. @riddhimore.bsky.social
March 7, 2025 at 3:22 PM
Looking for evidence of data leakage in the #HumanEval code generation #LLM benchmark? Check out our analysis of a subset of HumanEval tasks vs comparable tasks on #ChatGPT, #Claude and #Llama. @riddhimore.bsky.social
Grok vs DeepSeek
Grok = live data + trends + contextual analysis
DeepSeek = coding + structured reasoning + efficiency
Different strengths. Different workflows.
Explore the full comparison: aicomparison.ai/grok-vs-deep...
#GrokVsDeepSeek #Grok #DeepSeek #AI #AIAutomation
Grok = live data + trends + contextual analysis
DeepSeek = coding + structured reasoning + efficiency
Different strengths. Different workflows.
Explore the full comparison: aicomparison.ai/grok-vs-deep...
#GrokVsDeepSeek #Grok #DeepSeek #AI #AIAutomation
Grok vs DeepSeek 2026: Full AI Comparison & Benchmarks
Compare Grok and DeepSeek on AIME math scores, HumanEval coding accuracy, API pricing, and real-time data access before choosing an AI model in 2026.
aicomparison.ai
September 24, 2026 at 5:58 AM
Grok vs DeepSeek
Grok = live data + trends + contextual analysis
DeepSeek = coding + structured reasoning + efficiency
Different strengths. Different workflows.
Explore the full comparison: aicomparison.ai/grok-vs-deep...
#GrokVsDeepSeek #Grok #DeepSeek #AI #AIAutomation
Grok = live data + trends + contextual analysis
DeepSeek = coding + structured reasoning + efficiency
Different strengths. Different workflows.
Explore the full comparison: aicomparison.ai/grok-vs-deep...
#GrokVsDeepSeek #Grok #DeepSeek #AI #AIAutomation
Ask five frontier labs for their HumanEval scores and you’ll get nearly identical numbers. That’s not differentiation. It’s benchmark saturation.
See how labs are building fresh, expert-reviewed evaluations from production data: imerit.ai/resources/bl...
#LLMEvaluation #AIBenchmarks #GenAI
See how labs are building fresh, expert-reviewed evaluations from production data: imerit.ai/resources/bl...
#LLMEvaluation #AIBenchmarks #GenAI
September 25, 2026 at 3:36 PM
Ask five frontier labs for their HumanEval scores and you’ll get nearly identical numbers. That’s not differentiation. It’s benchmark saturation.
See how labs are building fresh, expert-reviewed evaluations from production data: imerit.ai/resources/bl...
#LLMEvaluation #AIBenchmarks #GenAI
See how labs are building fresh, expert-reviewed evaluations from production data: imerit.ai/resources/bl...
#LLMEvaluation #AIBenchmarks #GenAI
🆕 GPT-4o vs Claude 3.5 Sonnet: HumanEval Pass@1 Gap
#LLM #AI #CodeGeneration
https://tildalice.io/gpt-4o-vs-claude-3-5-sonnet-humaneval-benchmark/
#LLM #AI #CodeGeneration
https://tildalice.io/gpt-4o-vs-claude-3-5-sonnet-humaneval-benchmark/
June 1, 2026 at 6:04 PM
🆕 GPT-4o vs Claude 3.5 Sonnet: HumanEval Pass@1 Gap
#LLM #AI #CodeGeneration
https://tildalice.io/gpt-4o-vs-claude-3-5-sonnet-humaneval-benchmark/
#LLM #AI #CodeGeneration
https://tildalice.io/gpt-4o-vs-claude-3-5-sonnet-humaneval-benchmark/
Speaking of LLM’s and svelte 5, @khromov.se has built this to help
github.com/khromov/svel...
@sveltesociety.dev
github.com/khromov/svel...
@sveltesociety.dev
GitHub - khromov/svelte-bench: An LLM benchmark for Svelte 5 based on the HumanEval methodology
An LLM benchmark for Svelte 5 based on the HumanEval methodology - khromov/svelte-bench
github.com
May 9, 2025 at 2:46 PM
Speaking of LLM’s and svelte 5, @khromov.se has built this to help
github.com/khromov/svel...
@sveltesociety.dev
github.com/khromov/svel...
@sveltesociety.dev
Flash-dLLM speeds up diffusion LLM inference by tackling a hidden bottleneck: GPU memory I/O. With fused KV-cache kernels and self draft-and-verify decoding, it reaches 11× speedup on HumanEval over prior methods.
http://arxiv.org/abs/2609.26796v1
http://arxiv.org/abs/2609.26796v1
September 23, 2026 at 6:15 PM
Flash-dLLM speeds up diffusion LLM inference by tackling a hidden bottleneck: GPU memory I/O. With fused KV-cache kernels and self draft-and-verify decoding, it reaches 11× speedup on HumanEval over prior methods.
http://arxiv.org/abs/2609.26796v1
http://arxiv.org/abs/2609.26796v1
Watch and learn!
Let's observe Qwen2.5-coder:0.5b on OpenAI HumanEval.
`pip install observers`
And start collecting your data on the @huggingface.bsky.social Hub.
Dataset: huggingface.co/datasets/dav...
Library: github.com/cfahlgren1/o...
Let's observe Qwen2.5-coder:0.5b on OpenAI HumanEval.
`pip install observers`
And start collecting your data on the @huggingface.bsky.social Hub.
Dataset: huggingface.co/datasets/dav...
Library: github.com/cfahlgren1/o...
November 21, 2024 at 1:46 PM
Watch and learn!
Let's observe Qwen2.5-coder:0.5b on OpenAI HumanEval.
`pip install observers`
And start collecting your data on the @huggingface.bsky.social Hub.
Dataset: huggingface.co/datasets/dav...
Library: github.com/cfahlgren1/o...
Let's observe Qwen2.5-coder:0.5b on OpenAI HumanEval.
`pip install observers`
And start collecting your data on the @huggingface.bsky.social Hub.
Dataset: huggingface.co/datasets/dav...
Library: github.com/cfahlgren1/o...
Sub-3-bit quantization survives short answers but collapses on long reasoning. Distilling on the model's own quantized decoding path lifts BF16 retention from 35% to 70% on MATH-500 and 66% to 91% on HumanEval.
Low-bit cost shows up in long outputs.
Low-bit cost shows up in long outputs.
September 23, 2026 at 4:30 PM
Sub-3-bit quantization survives short answers but collapses on long reasoning. Distilling on the model's own quantized decoding path lifts BF16 retention from 35% to 70% on MATH-500 and 66% to 91% on HumanEval.
Low-bit cost shows up in long outputs.
Low-bit cost shows up in long outputs.
Style, Not Self: Surface Cues Explain Zero-Shot Code Attribution by Large Language Models
Ehsan Barkhordar, Surendrabikram Thapa
#arXiv #cs.AI #cs.CL #cs.SE
Ehsan Barkhordar, Surendrabikram Thapa
#arXiv #cs.AI #cs.CL #cs.SE
Style, Not Self: Surface Cues Explain Zero-Shot Code Attribution by Large Language Models
If a language model can recognize code it wrote, it may favor that code as a judge, and instances of one model monitoring each other could collude. We test this zero-shot on current commercial models. Five LLMs generate solutions to MBPP, HumanEval, and DS-1000, seven more to MBPP, and models act a…
arxiv.org
September 25, 2026 at 4:18 PM
Our paper "Addressing #DataLeakage in #HumanEval Using Combinatorial Test Design" has been accepted for publication at the Int’l Conference on Software Testing, Verification & Validation (#ICST2025), Short Papers, Vision & Emerging Results.
co-author: Riddhi More
arxiv.org/abs/2412.01526
#benchmark
co-author: Riddhi More
arxiv.org/abs/2412.01526
#benchmark
January 30, 2025 at 5:01 PM
Our paper "Addressing #DataLeakage in #HumanEval Using Combinatorial Test Design" has been accepted for publication at the Int’l Conference on Software Testing, Verification & Validation (#ICST2025), Short Papers, Vision & Emerging Results.
co-author: Riddhi More
arxiv.org/abs/2412.01526
#benchmark
co-author: Riddhi More
arxiv.org/abs/2412.01526
#benchmark
💡 For code tasks like MBPP and HumanEval, how you measure changes everything.
✅ Using continuous metrics (probability of correct answers) gives much better decision accuracy.
❌ Accuracy alone? No better than random 😬
✅ Using continuous metrics (probability of correct answers) gives much better decision accuracy.
❌ Accuracy alone? No better than random 😬
April 15, 2025 at 1:02 PM
💡 For code tasks like MBPP and HumanEval, how you measure changes everything.
✅ Using continuous metrics (probability of correct answers) gives much better decision accuracy.
❌ Accuracy alone? No better than random 😬
✅ Using continuous metrics (probability of correct answers) gives much better decision accuracy.
❌ Accuracy alone? No better than random 😬
Outside of HumanEval, the model pretrained on the SYNTH dataset outperforms the standard nanochat on every task. I report results for both the standard chat template (Harmony) as well as the Qwen3 Chat template used in the SYNTH dataset.
February 1, 2026 at 9:55 PM
Outside of HumanEval, the model pretrained on the SYNTH dataset outperforms the standard nanochat on every task. I report results for both the standard chat template (Harmony) as well as the Qwen3 Chat template used in the SYNTH dataset.
3/3
Current features:
• Python coder-reviewer loops
• HumanEval integration
• Custom prompts
• Configurable model combinations
• PDF reports
Current features:
• Python coder-reviewer loops
• HumanEval integration
• Custom prompts
• Configurable model combinations
• PDF reports
June 9, 2026 at 8:24 AM
3/3
Current features:
• Python coder-reviewer loops
• HumanEval integration
• Custom prompts
• Configurable model combinations
• PDF reports
Current features:
• Python coder-reviewer loops
• HumanEval integration
• Custom prompts
• Configurable model combinations
• PDF reports
AI benchmarks are the new vanity metrics.
"We scored 94.2% on HumanEval" tells you nothing about whether the model can debug a race condition in production at 3 AM.
Show me the benchmark for "handles ambiguity without hallucinating confidence."
"We scored 94.2% on HumanEval" tells you nothing about whether the model can debug a race condition in production at 3 AM.
Show me the benchmark for "handles ambiguity without hallucinating confidence."
March 29, 2026 at 7:11 AM
AI benchmarks are the new vanity metrics.
"We scored 94.2% on HumanEval" tells you nothing about whether the model can debug a race condition in production at 3 AM.
Show me the benchmark for "handles ambiguity without hallucinating confidence."
"We scored 94.2% on HumanEval" tells you nothing about whether the model can debug a race condition in production at 3 AM.
Show me the benchmark for "handles ambiguity without hallucinating confidence."
Qwen 3.8 27B spent 32,000 tokens and 14 minutes thinking about one HumanEval problem, then returned an empty response. It never reached an answer. That single problem was 7.2% of everything it generated across all 164.
#LocalAI
#LocalAI
Why Qwen 3.8 27B Feels Slow: Reasoning Tokens Measured
Qwen 3.8 27B generates at full speed on a 3090 and still crawls. All 164 HumanEval problems measured on the chat path: 92.8% of output is thinking.
insiderllm.com
August 17, 2026 at 11:00 PM
Qwen 3.8 27B spent 32,000 tokens and 14 minutes thinking about one HumanEval problem, then returned an empty response. It never reached an answer. That single problem was 7.2% of everything it generated across all 164.
#LocalAI
#LocalAI
A fun, clever idea from @upiter.bsky.social : treat code generation as a sequential editing problem -- this gives you loads of training data from synthetically editing existing code
And it works! Higher performance on HumanEval, MBPP, and CodeContests across small LMs like Gemma-2, Phi-3, Llama 3.1
And it works! Higher performance on HumanEval, MBPP, and CodeContests across small LMs like Gemma-2, Phi-3, Llama 3.1
Our paper showing that LMs benefit from human-like abstractions for code synthesis was accepted to ICLR! 🇸🇬
We show that order matters in code gen. -- casting code synthesis as a sequential edit problem by preprocessing examples in SFT data improves LM test-time scaling laws
We show that order matters in code gen. -- casting code synthesis as a sequential edit problem by preprocessing examples in SFT data improves LM test-time scaling laws
February 13, 2025 at 3:42 PM
A fun, clever idea from @upiter.bsky.social : treat code generation as a sequential editing problem -- this gives you loads of training data from synthetically editing existing code
And it works! Higher performance on HumanEval, MBPP, and CodeContests across small LMs like Gemma-2, Phi-3, Llama 3.1
And it works! Higher performance on HumanEval, MBPP, and CodeContests across small LMs like Gemma-2, Phi-3, Llama 3.1
trynimbus.org/nimbus-model... #buildinginpublic #buildinpublic my little 6gb 9B contribution to the open source community. Hoping for a solid MLX run so Mac Neo users can deploy it comfortably. Smaller and smarter…smaller and smarter… huggingface.co/Nimbus-Labs/...
Nimbus Models — Small models. Serious work.
Nimbus-9B v2.1 reaches 89.0% HumanEval and 87.3% MBPP with native thinking. See all four verified EvalPlus scores.
trynimbus.org
July 21, 2026 at 10:36 AM
trynimbus.org/nimbus-model... #buildinginpublic #buildinpublic my little 6gb 9B contribution to the open source community. Hoping for a solid MLX run so Mac Neo users can deploy it comfortably. Smaller and smarter…smaller and smarter… huggingface.co/Nimbus-Labs/...