Xiaomi’s MiMo AI team has open-sourced MiMo Code V0.1.0, a terminal-native AI coding assistant that the Chinese electronics giant says outperforms Anthropic’s Claude Code on key agentic coding benchmarks, especially on long-horizon, multi-step tasks (200+ steps) — at least, according to its own internal beta release and survey of 576 developers.

It’s also bundling limited-time free access to MiMo-V2.5, its multimodal flagship model with a million-token context window, requiring no registration to get started.

The release was announced June 10, 2026 in a post on the social network X from the official @XiaomiMiMo account, which described the tool as “more than an AI coding assistant in your terminal — it’s the smartest coding partner you’ll ever work with.”

MiMo Code is available now on GitHub under an MIT license, and installs with a single terminal command (curl -fsSL https://mimo.xiaomi.com/install | bash) on macOS and Linux or via npm (npm install -g @mimo-ai/cli) on Windows.

The project is a fork of the open-source OpenCode agent, which Xiaomi has extended with its own memory architecture, workflow modes, and model harness.

The end of AI coding agents’ amnesia?

As any avid vibe coder would surely attest, AI coding agents degrade over long working sessions: as the context window fills, earlier decisions, conventions, and task state get compacted away or lost entirely, forcing developers to re-explain their projects.

Xiaomi argues this approach is doomed at scale. “What we need is not better compression, but an explicit storage-and-retrieval mechanism that decides what information should be written into persistent structures, and when it should be recalled,” the MiMo team noted in their launch blog.

MiMo Code attacks this with a cross-session memory system, powered under the hood by SQLite FTS5 full-text search, that spans four layers: project memory (a persistent MEMORY.md file), session checkpoints, scratch notes, and per-task progress logs.

The note-taking is key, here: Rather than forcing the primary coding agent to pause its work to take notes, the system deploys an independent “checkpoint-writer” subagent.

Think of it the primary coding agent as a construction contractor working to build a massive mansion alongside a dedicated architect, the checkpoint-writer subagent. While the main agent focuses on building out the physical structure, the subagent updates the blueprints in real time, noting decisions, issues, and the actual lay of the land as the construction project progresses.

When the context window approaches its limits — the contractor gets lost in the half-built mansion — it can consult the subagent and find its place again. In the case of MiMo Code, the system simply rebuilds the environment from structured checkpoints with the relevant context, ensuring no loss of operational momentum.

Two self-improvement mechanisms round out the system: a /dream command that periodically (roughly every seven days) reviews historical sessions, deduplicates them, and compresses them into long-term memory, and a “distill” function that mines past sessions for repeated workflows that can be automated, following a similar approach taken recently by OpenAI and Anthropic with their various models.

Impressive performance on software engineering (SWE) benchmarks

According to benchmark figures published in Xiaomi’s technical blog post, MiMo Code paired with MiMo-V2.5-Pro outperformed Claude Code paired with Claude Sonnet 4.6 on all three evaluations tested:

• SWE-bench Verified: 82% vs. 79%

• SWE-bench Pro: 62% vs. 55%

• Terminal Bench 2: 73% vs. 69%

The harness itself accounts for a measurable share of the gain. Running the same MiMo-V2.5-Pro model in both harnesses, MiMo Code scored 62% on SWE-bench Pro versus 57% for Claude Code, and 73% on Terminal Bench 2 versus 68% — roughly five points each, attributable purely to the agent system rather than the model.

Xiaomi notably did not publish comparisons against OpenAI’s Codex or Google’s Gemini CLI — Claude Code is the sole named competitor throughout its materials, a telling choice of benchmark target.

Independent reference points suggest why. On the official Terminal-Bench 2.0 leaderboard maintained at tbench.ai, OpenAI’s Codex CLI running GPT-5.5 scores 82.2% — roughly nine points above MiMo Code’s self-reported 73% — and OpenAI’s own GPT-5.5 announcement claims 82.7% on the same benchmark.

On SWE-Bench Pro, however, the picture flips: OpenAI reports GPT-5.5 at 58.6%, below MiMo Code + MiMo-V2.5-Pro’s claimed 62%. (MiMo Code does not yet appear on either official leaderboard, and cross-comparing self-run numbers against leaderboard submissions carries the usual configuration caveats.)

Perhaps more interesting than the offline benchmarks: Xiaomi says it ran a human double-blind A/B evaluation during its internal beta, covering 576 developers working in 474 real private repositories, producing 1,213 judged head-to-head pairs against Claude Code using the same target model.

Under 200 execution steps, the two systems split roughly 50/50 — but past 200 steps, MiMo Code’s win rate rose above 65%, supporting the company’s thesis that its memory and state-management architecture pays off specifically on long-horizon work.

Xiaomi itself concedes the standard benchmarks “still measure one-shot problem-solving ability” and don’t capture the tool’s multi-session design goals.

As always, these are vendor self-reported numbers that haven’t been independently verified, and head-to-head harness comparisons are sensitive to configuration. But the claims are consistent with a broader industry pattern: scaffolding and harness engineering are becoming as important as raw model capability in agentic coding performance.

Easy integration with existing developer systems and voice control

From a user experience standpoint, MiMo Code is designed to live where developers already work. It operates directly in the terminal, reading and writing files, running commands, and managing Git.

Out of the box, the tool requires zero configuration, connecting automatically to “MiMo Auto”—a free-for-a-limited-time channel powered by Xiaomi’s multimodal MiMo V2.5 model, which boasts a massive million-token context window. For developers migrating from existing environments, the transition is frictionless: MiMo Code automatically imports MCP servers, custom skills, and API configurations from Claude Code.

Other noteworthy features include:

• Compose mode: Pressing Tab switches the agent into a specification-driven workflow in which the developer describes a high-level goal and the system autonomously executes the full development cycle — design, planning, coding, testing, and review — following what Xiaomi describes as a “heavy planning upfront, stable verification later” strategy.

• Voice control: Built on Xiaomi’s MiMo-ASR speech recognition with TenVAD voice activity detection, developers can dictate and modify instructions verbally and speak commands like “send” and “execute” for fully hands-free operation (available for logged-in users).

According to Xiaomi, the gains from the agent harness itself are measurable. Running the same underlying MiMo model in both harnesses, the company says MiMo Code scored 62% on SWE-Bench Pro versus 57% for Claude Code, and 73% on Terminal Bench 2 versus Claude Code’s 68% — roughly five percentage points better on each, attributable purely to the agent system rather than the model.

As always, these are vendor self-reported numbers that haven’t been independently verified, and head-to-head harness comparisons are sensitive to configuration. But the claim is consistent with a broader industry pattern: scaffolding and harness engineering are becoming as important as raw model capability in agentic coding performance.

Aggressively affordable

The bigger lure for many developers may be what’s bundled in.

MiMo Code ships with “MiMo Auto,” a zero-configuration channel offering free, limited-time access to MiMo-V2.5 — the natively multimodal model Xiaomi released in late April 2026, a sparse mixture-of-experts design with 310 billion total parameters (just 15 billion active per inference) and a 1 million token context window, which the company positions as matching Anthropic’s Claude Sonnet 4.6 in multimodal agentic work.

As VentureBeat reported when the MiMo-V2.5 family launched in April, the models are MIT-licensed and among the most efficient and affordable available for agentic tasks.

The larger MiMo-V2.5-Pro — a 1.02-trillion-parameter mixture-of-experts model with 42 billion active parameters and a hybrid-attention architecture — led the open-source field on Xiaomi’s ClawEval agentic benchmark with a 63.8% success rate while consuming only about 70,000 tokens per trajectory, roughly 40–60% fewer than Anthropic’s Claude Opus 4.6, Google’s Gemini 3.1 Pro, or OpenAI’s GPT-5.4 needed for comparable results.

Notably, the V2.5-Pro’s post-training was explicitly designed to instill “harness awareness” — training the model to manage its own memory and context within agent scaffolds like Claude Code or OpenCode — making a Xiaomi-built harness optimized around that capability a logical next step.

Pricing is similarly aggressive: MiMo-V2.5 starts at $0.40 per million input tokens and $2.00 per million output tokens, while V2.5-Pro runs $1.00/$3.00 per million (input/output) up to 256K context, doubling beyond that, with cache hits dropping input costs to as little as $0.20–$0.40 per million, making it among the cheapest frontier models available globally.

VentureBeat Frontier AI Model API Pricing Snapshot

For developers who don’t want Xiaomi’s models at all, MiMo Code also supports third-party backends — including token plans from DeepSeek, Moonshot’s Kimi, and Zhipu’s GLM — along with any OpenAI-compatible API, mirroring the bring-your-own-model flexibility of its OpenCode parent.

Terminal AI coding agent wars go global

MiMo Code lands in an increasingly crowded field of terminal-based coding agents: Anthropic’s Claude Code, OpenAI’s Codex CLI, Google’s Gemini CLI, and open-source players like OpenCode and Aider.

What’s new is the entrant. Xiaomi — the world’s third-largest smartphone maker, with a fast-growing EV business — has been methodically building its MiMo AI division since the release of the MiMo-7B reasoning model in April 2025, following with the MiMo-VL vision-language series, MiMo-V2-Flash, the 1-trillion-parameter MiMo-V2-Pro in March 2026, and the V2.5 flagship family in April.

The effort is led by Fuli Luo, a veteran of DeepSeek’s disruptive R1 project, who has characterized Xiaomi’s frontier push as a “quiet ambush” — and backed it with a 100-trillion free token grant for builders announced alongside the V2.5 launch.

The playbook is familiar from DeepSeek, Alibaba’s Qwen, MiniMax, and Moonshot AI’s Kimi series: release genuinely capable models and tooling under permissive licenses at a fraction of U.S. lab pricing, and convert the resulting developer mindshare into a durable ecosystem.

By pairing an open-source agent harness with a free frontier-class model, Xiaomi is effectively eliminating both the licensing and the usage cost of entry — at least for now.

What it means for enterprises and technical decision-makers

For engineering leaders, MiMo Code is a low-risk, potentially high-value evaluation candidate: MIT-style licensing permits modification and commercial integration, the OpenCode lineage means the architecture is inspectable, and the bring-your-own-model support means it can be pointed at an internally approved endpoint rather than Xiaomi’s cloud.

The persistent memory system addresses a real and widely felt pain point in agentic development workflows — one that competitors are also racing to solve.

The countervailing considerations: the “free for a limited time” model access is by definition temporary and routes code context through Xiaomi’s servers, which will be a non-starter for organizations with strict data-residency or IP policies; the benchmark edge over Claude Code is self-reported; and a V0.1.0 release number signals exactly what it suggests about maturity.

Teams subject to U.S. government procurement restrictions on Chinese technology vendors should also weigh that context before adopting.

GenAI image generators like Stable Diffusion do not draw a picture pixel by pixel from left to right. They start with noise and iteratively refine the entire image in parallel until it converges, in a process known as diffusion. For years, applying that same principle to text generation had remained out of reach at scale.

Standard language models work like a typewriter: one token at a time, left to right, with no ability to revise a committed output. That pattern works in the cloud, where batch sizes keep GPUs saturated. For local inference or low-concurrency deployments, the GPU is idle most of the time.

Google’s DiffusionGemma, released this week, is an open source experimental model that applies diffusion to text generation at production scale. Built on the Gemma 4 backbone and released under the Apache 2.0 license, it is the first diffusion language model natively supported in the open source vLLM inference platform. It generates a 256-token block in parallel rather than sequentially, with every token position attending to every other. Google says DiffusionGemma generates text up to 4x faster than standard models on GPUs. At batch size 1 on a single Nvidia H100, the FP8 version reaches 1,008 tokens per second. On H200, it hits 1,288 — roughly six times a standard autoregressive baseline, according to vLLM benchmark results published today.

Despite the speed gains, Google did not oversell the release. The company’s launch post acknowledged directly that DiffusionGemma’s overall output quality is lower than standard Gemma 4, adding “For applications that demand maximum quality, we recommend deploying standard Gemma 4.”

What DiffusionGemma does

DiffusionGemma does not generate tokens in order. It starts with a block of 256 random placeholder tokens, effectively a blank canvas, and runs multiple refinement passes over the entire block at once. On each pass, it evaluates every position and locks in the ones it is most confident about. Uncertain positions get randomized and reconsidered on the next pass, with the model using what it resolved in the previous round to inform the next attempt. The block converges progressively until enough positions stabilize to anchor the rest.

Two things follow from that architecture.

• Self-correction. An autoregressive model that commits to a wrong token is stuck with it, because subsequent tokens are already conditioned on the mistake. DiffusionGemma can identify low-confidence positions and re-evaluate them on the next pass.

• Bidirectional context. Every position attends to every other position in the block simultaneously, including tokens that appear later in the sequence. That makes the model structurally better suited to constrained generation tasks where left-to-right generation fails.

Google demonstrated both properties with a fine-tuned Sudoku solver. The base model solved zero puzzles. After fine-tuning on a Sudoku dataset, it reached an 80% success rate and converged in 12 denoising steps rather than 48. The efficiency gain came directly from the model’s ability to self-correct and stop early.

How it was built

DiffusionGemma runs as a 26B Mixture of Experts model that activates only 3.8B parameters during inference. Quantized, it fits within 18GB VRAM on consumer hardware including the Nvidia RTX 4090 and 5090. Google and NVIDIA also optimized for enterprise Hopper and Blackwell servers using NVFP4 kernels.

The vLLM integration required new work because DiffusionGemma does not fit the standard serving model. A typical vLLM batch applies the same attention type to every request. DiffusionGemma requests alternate between causal and bidirectional attention as they cycle through prompt reading, canvas refinement and block commit. The team built per-request attention switching into both the Triton and FlashAttention 4 backends and reused the existing speculative decoding path for the refinement loop.

The new ModelState interface the team built for this integration is designed to support additional diffusion models in vLLM as they emerge.

Where the speed wins and where it does not

DiffusionGemma’s speed advantage is real but conditional. Where it applies depends entirely on deployment context.

The numbers. At batch size 1 on a single H100, vLLM’s published benchmarks put the FP8 model at roughly five times a standard autoregressive baseline. On H200, roughly six times. Those peak figures reflect optimal conditions: single user, dedicated hardware, FP8 quantization.

Where it wins. Local inference, single-user applications and low-concurrency serving. In those conditions the GPU has spare compute and memory bandwidth is the bottleneck. DiffusionGemma’s parallel block generation fills that gap.

Where it does not. High-throughput cloud serving. When a server is batching hundreds of concurrent requests, autoregressive models already saturate available compute and DiffusionGemma’s parallel decoding provides diminishing returns.

The quality ceiling. Guilherme O’Tina, an AI researcher, put a finer point on it on X. “Local artifacts vs hallucinations are different problems and that decides where this actually wins,” O’Tina wrote.

How it compares

Diffusion language models are not new. Researchers have built them at smaller scales for several years, and Inception Labs’ Mercury Coder applied the approach commercially to coding tasks in 2025. What DiffusionGemma adds is scale — a 26B MoE backbone, native vLLM serving and a general-purpose instruction-tuned model rather than a domain-specific one.

The more useful comparison for engineers evaluating this against existing inference tooling is speculative decoding, and the distinction matters. Speculative decoding keeps a standard autoregressive target model and uses a smaller draft model to guess several tokens ahead. The target model verifies them in one pass. If sampling is correct, the output distribution stays identical to the target. The architecture is unchanged.

Andrew Kuncevich, an ML and AI researcher focused on production AI systems, put it directly on X. “DiffusionGemma is different. It does not just guess future tokens. It creates a noisy 256-token canvas and repeatedly denoises the whole block in parallel. So it’s not just a decoding trick — it’s a different generation paradigm,” Kuncevich wrote.

Compared to standard Gemma 4, the trade is speed for quality. Google’s benchmark data shows DiffusionGemma below standard Gemma 4 on general output quality metrics, with the gap varying by task.

On structured constrained tasks, including code infilling, template generation and problems requiring bidirectional constraint propagation, the architecture has a structural advantage that fine-tuning can surface, as the Sudoku result demonstrates. On open-ended generation, standard Gemma 4 remains the stronger option.

What this means for enterprises

DiffusionGemma serves via a standard vLLM OpenAI-compatible endpoint with no diffusion-specific pipeline changes required.

This is not a general-purpose model upgrade.

For teams running local or low-concurrency inference, the architecture choice just expanded. Until now, cutting generation latency on dedicated GPU hardware meant using a smaller model and accepting the quality trade-off. DiffusionGemma offers a third path at the same parameter footprint, on consumer hardware, with same-day vLLM support.

For constrained generation workloads, bidirectional attention is worth evaluating. Code infilling, structured data generation and tasks where correct output depends on context not yet generated are where this architecture has a structural edge.

The ModelState interface built for this integration is designed to generalize as additional diffusion models emerge.

The quality trade-off is real and Google acknowledges it. For teams running local inference on dedicated GPU hardware, this is worth testing.