LIVE
Introducing Cosmos 3 Edge20/07/26 · Hugging Face|Kimi K3, Qwen 3.8, and Anthropic's (Potential) Unravelling20/07/26 · Anthropic|At SIGGRAPH, NVIDIA Advances Graphics and Simulation With Agentic and Physical AI20/07/26 · NVIDIA|China's open-weights AI strategy is winning20/07/26|Bristol Myers Squibb Building Life Science Industry’s Most Advanced AI Factory on NVIDIA Vera Rubin20/07/26 · NVIDIA|Safety and alignment in an era of long-horizon models20/07/26 · OpenAI|Claude Fable produced a counterexample to the Jacobian Conjecture20/07/26 · Anthropic|Apply for Anthropic’s AI for Science rare disease research grants20/07/26 · Anthropic|AI advice made people less accurate but more confident – sudy19/07/26|Vincentwei1021/video-shotcraft: AI video skill for Claude Code & Codex — cinematic product videos with Remotion: 106 shot recipe cards, 161 motion previews, a production-ready template19/07/26 · Anthropic|Claude Code uses Bun written in Rust now19/07/26 · Anthropic|OpenAI reduces Codex Model Context Size from 372k to 272k19/07/26 · OpenAI|Introducing Cosmos 3 Edge20/07/26 · Hugging Face|Kimi K3, Qwen 3.8, and Anthropic's (Potential) Unravelling20/07/26 · Anthropic|At SIGGRAPH, NVIDIA Advances Graphics and Simulation With Agentic and Physical AI20/07/26 · NVIDIA|China's open-weights AI strategy is winning20/07/26|Bristol Myers Squibb Building Life Science Industry’s Most Advanced AI Factory on NVIDIA Vera Rubin20/07/26 · NVIDIA|Safety and alignment in an era of long-horizon models20/07/26 · OpenAI|Claude Fable produced a counterexample to the Jacobian Conjecture20/07/26 · Anthropic|Apply for Anthropic’s AI for Science rare disease research grants20/07/26 · Anthropic|AI advice made people less accurate but more confident – sudy19/07/26|Vincentwei1021/video-shotcraft: AI video skill for Claude Code & Codex — cinematic product videos with Remotion: 106 shot recipe cards, 161 motion previews, a production-ready template19/07/26 · Anthropic|Claude Code uses Bun written in Rust now19/07/26 · Anthropic|OpenAI reduces Codex Model Context Size from 372k to 272k19/07/26 · OpenAI|
🇪🇺 Europe (collective)💻 CodeReleased February 2024

StarCoder 2

BigCode
Open-sourceFreeSelf-hostRGPD
Verdict

A solid pick for self-hosting and full data control.

Overview

StarCoder 2 comes from the BigCode collective (HuggingFace + ServiceNow). Apache 2, fully transparent training data, ideal for absolute sovereignty or local/edge deployments. Performance trailing frontiers but it's the only fully transparent one.

Skill profile

FrenchReasoningSpeedCreativitySafety
Verified benchmarks
HumanEval46.3%

What do these scores mean?

HumanEval46.3%fair

Python code generation from specifications.

Strengths

  • Has a free tier
  • Open-source and self-hostable

Limitations

  • Light safety filters
  • API pricing not disclosed

Who is it for

A good fit if…
  • you build with a coding agent
  • you want to control cost or self-host
  • you have GDPR constraints
Skip it if…
  • you want strict guardrails

Ideal use cases

  • Local autocompletion
  • Code on edge
  • Lightweight apps
  • IDE plugins

Access & availability

Free tier availableSelf-hostable (open weights)

Key specifications

Context
16K
Input price
Output price
Speed
200 tok/s
Price not auditedBenchmarks not audited

Privacy

GDPR-compliantHosting : EU/selfAnonymizable data

Run it locally

Desktop with solid GPU / 32GB+ Mac11k674
QuantizationDiskRAM / VRAMTypical hardware
Q4 · recommended9.1 GB12 GB32GB PC / 24GB Mac / RTX 4070 Ti+
Q8 · balanced17.1 GB21 GB32GB Mac / RTX 3090-4090
FP16 · max quality32 GB37 GB64GB Mac / dual 24GB GPUs

Estimates for a moderate context. Long contexts need more RAM (KV cache).

🤗 Hugging Face
ollama run starcoder2:15b
Advanced data · for experts
Architecture, modalities, detailed cost, full benchmarks
Architecture
Dense
Size
15B
Cutoff date
Aug 2023
Inputs
text
Outputs
code
License
apache-2
Hosting
EU/self
Hallucination score
2/5
Estimated cost (API)
Typical exchange (~3k in / 1k out)
Not disclosed
1M in + 1M out
Not disclosed
All benchmarks
Arena Elo
MMLU
GPQA
HumanEval46.3%
SWE-Bench
MATH

Similar models