[AINews] Building AIE CODE: The top Coding Agent Conference

What it claims

This AINews issue (Latent Space, covering 28 to 29 October 2026) carries practical detail on Mistral Large 4 ("Le Chonk"). Mistral Large 4 is an open-weights mixture-of-experts model with 1.05T total and 52B active parameters, 1M context and text+image input. It is priced at $0.70/$2.10 per 1M input/output tokens ($0.07 cached). Its MEOW licence requires a separate commercial licence for model-as-a-service businesses. The issue also reports a supply-chain compromise of the Claude Code VS Code extension and changes to the Claude API and Codex limits. The AIE CODE section is an invite-only, single-track summit plug: about 1,000 seats, a couple hundred left, and code LS30 for 30% off for paid subscribers.

How it works

A community analysis by @eliebakouch describes the Mistral Large 4 architecture. It uses full sparse attention in the style of DeepSeek V3.2 and GLM 5. The KV cache is shared across NoPE layers after layer 35, and indexers are shared on RoPE layers from layer 17. It adds SSMax per-head logit scaling and LeRoPE (YaRN plus a learned per-frequency term), with no KV compression.

The Hugging Face card, as quoted on Reddit, gives further detail:

• Experts: 128 routed experts plus 1 shared expert, with 4 routed experts active per token.
• Features: configurable reasoning effort (three modes), function calling and strict JSON schema.
• Training: trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs.

Evidence

Artificial Analysis (AA) scores Mistral Large 4 at 26 on Intelligence Index v5. That is 1 point above Muse Spark Open and 6 points behind DeepSeek V4.1 Flash 1017 and GLM-5.3.

Coding results are weaker:

• FrontierCode 2.0: 15%, versus 29% for DeepSeek V4.1 Flash 1017 and 22% for GLM-5.3.
• Terminal-Bench 4.0: 25%, versus 8% for Muse Spark Open.
• Coding Agent Index: 48 at 8.5M tokens per task, versus 49 at 15.9M tokens for GLM-5.3.

It uses 82K output tokens per Intelligence Index task, 39% fewer than the Preview, at $2.06 per task. The AA-Omniscience hallucination rate fell from 41% to 28%, but accuracy is 26%. Vals ranks it #5 among open-weight models and #17 overall. Mistral's own claims include #1 among open models on FrontierCode and a 64% human-preference win rate against GLM 5.3 and Kimi K3.

How to use it

The Hugging Face repository is mistralai/Mistral-Large-4-1T-A52B, and the API model name is mistral-large-4.

• vLLM: supports the FP8 and NVFP4 checkpoints with a built-in DSpark drafter, giving 1.92 to 2.23x output throughput.
• llama.cpp: PR #31376 is merged. GGUF builds exist from ggml-org (including NVFP4) and Unsloth.
• Local: Unsloth's 1-bit GGUF is 223GB, versus 1.1TB in FP8. It keeps 91% top-1 accuracy and runs on a Mac Studio.
• Fine-tuning: TRL runs single-node LoRA SFT on 8×B300 at 142 GiB per GPU, about 8 seconds per 32.7K-token step.
• RL: SkyRL with vLLM ran LoRA RL on NVFP4 across 8 GB200 nodes. AIME 2024 rose from about 64% to about 75% in 75 steps.

Other actionable items:

• Claude API: maximum request size rose from 32MB to 128MB. The per-image limit rose to 3000px for requests with more than 20 images.
• Codex: usage limits doubled for Pro 100/200/500 and Business Premium.
• Google Jules: being wound down and folded into Antigravity.
• safeagents: Hugging Face's open proxy blocks out-of-scope agent tool calls before execution. Examples include metadata-endpoint requests, credential exfiltration and verifier tampering.

Limits

• Claude Code extension: @mattjay reports that the official Claude Code VS Code extension was compromised. Installing it during a brief window would have added an SSH backdoor. Anyone who updated recently should check.
• Claude Code quotas: a Reddit user reports that Anthropic support said retries from service errors count toward Claude Code limits and will not be compensated. Another user measured output per 1% of quota falling from 119K to 31K tokens week over week on Opus 5.5. That user attributes the drop to harness changes, such as subagents rereading parent context, with degradation above about 350K tokens of subagent context. Both are user reports.
• Mistral Large 4 licence and fit: model-as-a-service use needs a separate licence. Commenters say it is not competitive locally against DeepSeek V4.1, and one estimates it is undertrained given Mistral's roughly 4K-GPU footprint.
• Unverified claims: Gemini 4 Carbon benchmarks lack methodology, and the model is not publicly available. Qwen4 (Max, Flash, Plus, 27B, 1M context) is a preview with no specs. The Opus 5.5 math and ETH claims are unverified.