LLMWatch

87

BenchLM
open leader

80.6%

SWE-bench
Verified

1M

Baseline
Context

$0.18

Cheapest
per M tok

Frontier coding/agentic performance with native multimodal capabilities (image and video inputs). Native 1M-token context window.

The open-weight leaderboard champion. Delivers a BenchLM score of 87 and SWE-bench Verified score of 80.6% with native 1M-token context.

High-throughput open-weight model with a BenchLM score of 85. Designed for RAG and coding tasks with excellent speed ergonomics.

Strong reasoning-focused open model scoring 83 on BenchLM. Integrates a 203K-token context window optimized for complex math and logic.

Laptop-friendly local multimodal flagship from Google. Integrates native audio and image inputs directly into the LLM backbone without external encoders.

Alibaba's new generation. Max represents the current open-source quality leader on WhatLLM, paired with a solid mid-tier Plus variant.

Agentic coding wins

The strongest releases now optimize for coding agents, browser-use, tool-use, and terminal workflows rather than pure benchmark breadth.

MoE stays dominant

Sparse and hybrid MoE designs remain the best route to frontier capability without explosive serving costs (e.g. DeepSeek V4 Pro, Kimi K2.6).

Chinese labs lead pace

DeepSeek, Qwen, GLM/Z.AI, MiniMax, and Xiaomi continue to set much of the open-weight release tempo.

Licensing drives adoption

Apache 2.0 and MIT remain the most deployment-friendly licenses. Buyers are increasingly selecting models by legal fit as much as score.

Speed is a weapon

High-throughput models like MiniMax-M2.7 and Kimi K2.6 show that latency and tokens/sec are now first-class product metrics.

Frontier Coding (self-hosted)

DeepSeek V4 Pro → Kimi K2.6

DeepSeek is the safest benchmark anchor for self-hosting; Kimi K2.6 brings strong throughput and a very competitive composite score.

Cost-Efficient Production

MiMo-V2.5 → DeepSeek V4 Flash

MiMo-V2.5 is the cheapest high-ranking open option ($0.18/M); DeepSeek V4 Flash offers a stronger benchmark anchor for slightly higher budgets.

Single-GPU Coding

Devstral Small 2 → Mellum2

Devstral Small 2 (68.0% SWE-Verified) targets local consumer hardware; Mellum2 is optimized for low-latency coding workflows.

Computer Use Agents

MiniMax M3 → Step 3.7 Flash

MiniMax M3 is the standout open-weight candidate for agentic tasks; Step 3.7 Flash is the practical fallback with strong tool benchmarks.

MiniMax M3 Weights

Announced on June 1. Check HuggingFace for weights and final license terms.

Qwen Open-Weight Cadence

Qwen3.7 Max & Plus are live; look out for new open-weight coding variants.

DeepSeek V4 Family Stabilization

V4 Pro and Flash aim to establish a new price/performance ceiling for open-weight models.

Mistral's Coding Cadence

Devstral 2 and Small 2 continue Mistral's strong orientation toward developer-first self-hosting.