Model market

ByteDance Seedream 4.0

ByteDance's mainline text-to-image model — reliable prompt following and clean text rendering, billed per image.

ByteDance Seedream 4.5

An upgraded Seedream tier with finer detail and stronger layout control than 4.0.

Kimi K3

Moonshot AI's flagship — reasoning with native vision understanding and a 1M-token context window.

Qwen3 235B A22B Instruct 2507

A 235B MoE (22B active) instruct model from the July 2025 Qwen3 refresh.

Qwen3 Coder 480B A35B Instruct

Qwen's largest coding model — a 480B MoE (35B active) built for repo-scale codegen and agents.

Qwen3.5 397B A17B

The big open Qwen3.5 MoE — 397B parameters total, 17B active per token.

Qwen3.5 Plus

Alibaba's hosted mainline Qwen3.5 tier — balanced quality, speed and price.

ByteDance Seedream 3.0

The earlier Seedream generation — a budget text-to-image option that still renders text well.

ByteDance Seedream 5.0 Lite

The light tier of the newest Seedream line — fast, low-cost images for high-volume use.

DeepSeek Prover V2 671B

A 671B model specialized in formal theorem proving rather than general chat.

DeepSeek R1 (0528)

The 0528 refresh of DeepSeek's R1 reasoning model — long chain-of-thought answers for hard problems.

DeepSeek R1 Distill Llama 70B

R1 reasoning distilled into a 70B Llama base — much cheaper reasoning at some quality cost.

DeepSeek R1 Turbo

An R1 variant on faster serving, tuned for lower-latency reasoning workloads.

DeepSeek V3 (0324)

The 0324 snapshot of DeepSeek V3, the general-purpose MoE chat model.

DeepSeek V3 Turbo

DeepSeek V3 on faster serving — lower latency for everyday chat and tool use.

DeepSeek V3.1

V3.1 merges V3's chat strengths with R1-style thinking in one hybrid model.

DeepSeek V3.1 Terminus

The final V3.1 build, with tool-use and agent fixes, before the V3.2 line.

DeepSeek V3.2

DeepSeek V3.2 — sparse attention makes long-context inference markedly cheaper.

DeepSeek V3.2 Exp

The experimental V3.2 build that introduced DeepSeek Sparse Attention, priced aggressively low.

DeepSeek V4 Flash

The fast, low-cost tier of the V4 line — the default pick for high-volume chat.

DeepSeek V4 Pro

DeepSeek's current flagship — the strongest general model of the V4 line.

Doubao-Seedance 2.0 (Image-to-Video)

Animates a still frame into a short clip with ByteDance's Seedance 2.0.

Doubao-Seedance 2.0 (Text-to-Video)

Text-to-video with Seedance 2.0 — short clips straight from a prompt.

Doubao-Seedance 2.0 Fast (Image-to-Video)

The fast Seedance tier — quicker, cheaper image-to-video for drafts and volume work.

GLM-4.5

Zhipu's agent-focused flagship of the 4.5 generation — chat, reasoning and tool use in one model.

GLM-4.5-Air

The light GLM-4.5 tier — most of the capability at a fraction of the price.

GLM-4.5V

The vision variant of GLM-4.5 — screenshots, documents and visual grounding.

GLM-4.6

GLM-4.6 extends 4.5 with a longer context window and stronger coding.

GLM-4.6V

The vision build of the GLM-4.6 generation.

GLM-4.7

The last 4.x generation before GLM-5 — a further coding and agent refinement.

GLM-5

Zhipu's fifth-generation flagship — a step up in reasoning and agentic coding over 4.x.

GLM-5.1

A GLM-5 line refresh — incremental quality gains over the base model.

GLM-5.2

The current head of the GLM-5 line.

Kimi K2 0905

The September 2025 snapshot of K2, Moonshot's trillion-parameter open MoE for agentic work.

Kimi K2 Instruct

The instruction-tuned K2 base build — general chat and tool calling.

Kimi K2 Thinking

K2 with explicit long-horizon thinking — strong on multi-step agent and search tasks.

Kimi K2.5

A K2-line refresh that sharpens agentic tool use over the original K2.

Kimi K2.6

The latest K2-line refresh before K3 — incremental gains over K2.5.

MiniMax M1 80k

MiniMax's earlier reasoning MoE, with an 80k-token thinking budget.

MiniMax M2

A compact, fast MoE built for coding and agent loops — MiniMax's price-performance play.

MiniMax M2.1

An M2-line refresh — better coding and agent reliability over M2.

MiniMax M2.5

A further M2-line upgrade, balancing speed against quality.

MiniMax M2.7

The newest M2-line model — MiniMax's current mainline chat tier.

MiniMax-H3 (Image-to-Video)

Animate a still first frame into a 2K clip with synchronized audio — MiniMax's flagship image-to-video.

MiniMax-H3 (Reference-to-Video)

Keep a subject identity consistent across a 2K clip from one reference image.

MiniMax-H3 (Text-to-Video)

Flagship text-to-video with synchronized audio, 2K output, and 4–15s clips.

Qwen2.5 72B Instruct

The dense 72B workhorse of the earlier Qwen2.5 generation — stable and widely deployed.

Qwen3 Coder 30B A3B Instruct

A light coding MoE (30B total, 3B active) for cheap autocomplete and small agents.

Qwen3 Coder Next

The rolling "next" build of Qwen's coder line — the newest coding capabilities as they land.

Qwen3 VL 30B A3B Instruct

A compact vision-language MoE that reads images, UI screenshots and documents.

Qwen3.5 122B A10B

A mid-size Qwen3.5 MoE (122B total, 10B active) — near-flagship quality at lower cost.

Qwen3.5 27B

A dense 27B Qwen3.5 — small enough to be cheap, big enough to be useful.

Qwen3.5 35B A3B

The efficiency tier of Qwen3.5 — a 35B MoE with 3B active, made for high-throughput serving.

Vidu Q3 Pro (Image-to-Video)

Image-to-video on Vidu's Q3 Pro tier — animate one still into a clip.

Vidu Q3 Pro (Start-End-to-Video)

Give a start and an end frame and Q3 Pro interpolates the motion between them.