MiMo (Xiaomi)
A Tier · 8.3/10
Xiaomi's MiMo family. MiMo-V2.6 (2026-09-22) open-sources Pro (1.02T total / 42B active MoE) and Flash (309B / 15B active) under MIT, both natively omnimodal (text, image, video, audio) with 1M context. API prices unchanged from V2.5: Pro $0.435 input / $0.87 output per 1M tokens, Flash $0.14 / $0.28. MiMo-V2.5 TTS and open-source ASR round out the voice stack
Score Breakdown
Personality & Tone
Xiaomi's voice-first agentic stack
Tone: Direct, multimodal-aware. MiMo-V2.5-Pro is comfortable mixing image, audio, and text inputs in a single turn -- it's been trained for that, not retrofitted to it.
Quirks: Voice-pipeline orientation makes MiMo unusually expressive when audio is in the loop -- TTS variants (VoiceDesign, VoiceClone) and ASR are surfaced as first-class products, which most Chinese frontier vendors haven't done. PRC content filters apply on chat surfaces.
The Good and the Bad
What we like
- +Open weights at the flagship tier: MiMo-V2.6-Pro and Flash ship under MIT on Hugging Face, with the technical report, training environments and RL code also released
- +Vendor-reported agent scores put V2.6-Pro close to closed frontier models: DeepSWE v1.1 71.9, Terminal Bench 2.1 89.9, OSWorld-Verified 82.0 and Toolathlon-Verified 76.9, against 74.0, 89.1, 83.4 and 80.6 for Claude Opus 5 in the same table
- +Full voice pipeline shipped together: a frontier reasoning model (Pro), a multimodal base, a TTS family, and an open-source ASR -- Xiaomi positions MiMo-V2.5 as 'voice for the agent era,' which is rare in 2026 (most vendors ship one of these and integrate the others later)
- +Native omnimodal input in MiMo-V2.5-Pro and both V2.6 models is the differentiator: text, image, video and audio in one model, not bolted on after the fact. Closer to the Gemini 2.5 / GPT-5.5 design than to text-first models with separate vision adapters
- +Removing the surcharge for the full 1M-context tier at launch is a real value move -- Alibaba, Anthropic, and OpenAI all charge meaningfully more per token for full-context windows. Xiaomi flattening this lowers the barrier to long-document and agentic workloads
- +Open-source MiMo-V2.5-ASR is the practical takeaway for privacy-sensitive teams. Cohere Transcribe + Whisper had been the open-ASR options through 2025; MiMo-V2.5-ASR adds a Chinese-dialect-strong third entry
- +Xiaomi cites an Artificial Analysis Intelligence Index score of 46.32 for MiMo-V2.6-Pro (v4.3, September 2026), which it calls the highest of any open-source model
What could be better
- −Xiaomi's own table shows clear gaps on harder terminal and exploit tasks: Terminal Bench 4.0 34.9 and ExploitBench 47.9 for V2.6-Pro, against 49.0 and 70.0 for Claude Opus 5
- −Self-hosting Pro needs a large multi-GPU setup: Xiaomi's vLLM example uses 8-way tensor parallel and its SGLang example runs 2 nodes at 16-way. Flash is lighter, with 4-way (vLLM) and 8-way (SGLang) tensor-parallel examples. The cards do not state GPU models or VRAM
- −Most V2.6 benchmark numbers come from Xiaomi, and several are in-house suites (MiMo Code Bench, MiMo Visual Coding, MiMo Cyber Bench), which warrants the usual self-reporting discount
- −PRC content filters apply on Pro and Multimodal -- the same regulated-topic refusals that Hy3, DeepSeek, and Qwen exhibit. ASR is less affected by content filtering since it's transcription, not generation
- −English creative-writing polish lags Western frontier models -- pick MiMo for Chinese-language work, multimodal reasoning, or voice pipelines first, English prose second
- −Geo-availability: API access for non-Chinese developers may require an Xiaomi developer account and KYC; check the docs before assuming OpenAI-style account-creation friction
Pricing
Free (consumer)
- ✓Xiaomi consumer device integration (HyperOS, Mi AI)
- ✓Web chat at mimo.xiaomi.com
- ✓Basic usage limits apply
API (MiMo-V2.6-Pro)
- ✓1.02T total / 42B active MoE; text, image, video and audio input; 1M context
- ✓Cache-hit input $0.0036 per 1M tokens; cache writes free for a limited time
- ✓One price across the full 1M context in Xiaomi's V2.6 price table
- ✓Pro-UltraSpeed mode: up to 20x output speed at $4.35 input / $8.7 output per 1M tokens
- ✓OpenAI- and Anthropic-API-compatible endpoints (the standard pattern Chinese frontier models adopted in 2025-26)
- ✓Token Plan available for predictable high-volume use
API (MiMo-V2.6-Flash)
- ✓309B total / 15B active MoE with the same omnimodal input and 1M context as Pro
- ✓Cache-hit input $0.0028 per 1M tokens
- ✓About a third of Pro's price; vendor-reported DeepSWE v1.1 67.9 vs 71.9 for Pro
Self-hosted MiMo-V2.6 (MIT open weights)
- ✓Pro and Flash weights on Hugging Face and ModelScope, as RL and MOPD checkpoints
- ✓MOPD checkpoints (2026-09-27) mitigate tool-call repetition in agent loops
- ✓MiMo-V2.6-Distill-Qwen-9B: a 9B SFT checkpoint on Qwen3.5-9B for agentic RL research
- ✓Technical report, training environments and RL code released
MiMo-V2.5-TTS (3 sub-models)
- ✓Base TTS (general voice synthesis)
- ✓VoiceDesign (designed-from-scratch synthetic voices)
- ✓VoiceClone (replicate a target voice from a sample)
MiMo-V2.5-ASR (open-source)
- ✓Open-source under the MIT licence (Hugging Face, 2026-04-23)
- ✓English + Mandarin Chinese + major Chinese dialects (Cantonese, Shanghainese, etc.)
- ✓Self-hostable for privacy-sensitive transcription workloads
System Requirements
Hardware needed to self-host. Min = smallest viable setup (usually heavy quantization). Max = full-precision / production-grade.
| Model variant | Min | Max |
|---|---|---|
| MiMo-V2.6-Pro (1.02T total / 42B active MoE)MIT open weights (RL and MOPD checkpoints). Text, image, video and audio input, 1M context. The card does not state GPU model or VRAM | vLLM example: 8-way tensor parallel | SGLang example: 2 nodes, 16-way tensor parallel |
| MiMo-V2.6-Flash (309B total / 15B active MoE)MIT open weights (RL and MOPD checkpoints). The card does not state GPU model or VRAM | vLLM example: 4-way tensor parallel | SGLang example: 8-way tensor parallel |
| MiMo-V2.5-ASR (open-source)MIT licence on Hugging Face. Self-hostable for privacy-sensitive transcription. Strong specifically on Chinese dialects (Cantonese, Shanghainese) | 8 GB VRAM (RTX 3060 tier) for English + standard Mandarin | 1× A100 40 GB for full dialect coverage at production throughput |
Known Issues
- VERSION BUMP (2026-09-22): Xiaomi released and open-sourced the MiMo-V2.6 series under MIT (Hugging Face repos created 2026-09-21): Pro (1.02T total / 42B active MoE) and Flash (309B / 15B active), both natively omnimodal with 1M context. API pricing is unchanged from V2.5: Pro $0.435 input / $0.87 output, Flash $0.14 / $0.28 per 1M tokens, plus Pro-UltraSpeed at up to 20x output speed for $4.35 / $8.7. MiMo Desktop left early access with both models built in. Correction: this page previously called MiMo-V2.5-Pro API-only, but its weights have been on Hugging Face under MIT since 2026-04-27Source: Xiaomi MiMo blog (mimo.xiaomi.com/mimo-v2-6), Hugging Face XiaomiMiMo model cards and model list · 2026-09-22
- Tool-call repetition: Xiaomi says that after the V2.6 release, repeated identical or near-identical tool calls became one of the most noticeable issues in agentic use, consuming time and context without progress. The MiMo-V2.6-Pro-MOPD and Flash-MOPD checkpoints (2026-09-27) mitigate it, so self-hosters running agents should use them over the RL checkpointsSource: Hugging Face (XiaomiMiMo/MiMo-V2.6-Pro-MOPD card) · 2026-09-27
- OLDER SERIES RETIRED (2026-06-30): Xiaomi fully deprecated the prior **MiMo-V2 series** (V2-Pro / V2-Omni / V2-Flash) on 2026-06-30 -- migrate to the MiMo-V2.5 family this page covers, which is the designated replacement. If you pinned a `mimo-v2-*` model id, it's now failing; the V2.5 line is current. (Reinforces V2.5 as the live family; no change to our review scope.)Source: Xiaomi MiMo docs (mimo.mi.com/docs updates), 36Kr, Hacker News · 2026-06-30
- NEW PRODUCT (2026-06-11): **MiMo Code v0.1.0** -- Xiaomi's open-source terminal AI coding agent (MIT license, TypeScript, built on OpenCode tech), 4,300+ GitHub stars on launch day and a top Hacker News story (357 points). MiMo-V2.5 model built in FREE; also supports DeepSeek and Kimi models. Differentiators: persistent memory system (project memory + session checkpoints + task progress), a '/dream' weekly self-maintenance agent, Compose mode (idea → design/plan/code/test/review pipeline), voice control, and curl/npm install (`mimo` command, Windows via npm). Vendor benchmark claims (UNVERIFIED third-party): SWE-Bench Pro 62% vs Claude Code 57%, Terminal-Bench 73% vs 68%. It's a v0.1.0 -- expect churn -- but a free, open-source, frontier-backed CLI agent is direct pressure on Claude Code / Codex CLI / Grok Build pricing. Standalone review page candidate once it stabilizes past 0.xSource: GitHub (github.com/XiaomiMiMo/MiMo-Code), mimo.xiaomi.com/mimocode, Hacker News, Gizmochina · 2026-06-11
- MIMO CODE DAY-2 CAVEATS (2026-06-12, press analysis): the launch benchmarks deserve skepticism -- they're self-reported, and the 'beats Claude Code' comparison pits MiMo Code + MiMo-V2.5 against Claude Code running **Sonnet 4.6**, not the flagship models. VentureBeat's hands-on found genuine strength on ultra-long 200+ step tasks. Privacy note: the 'free for a limited time' model routes your code through Xiaomi servers -- a data-residency consideration for commercial codebases. v0.1.0 remains the only release (no hotfixes yet)Source: TechTimes (2026-06-12), VentureBeat (2026-06-11), GitHub releases · 2026-06-12
- MiMo-V2.5 family launched 2026-04-22 with four product lines released in parallel: Pro (1T/42B MoE, 1M context, native vision+audio), Multimodal base, TTS (Base + VoiceDesign + VoiceClone), and open-source ASR. This is Xiaomi's first explicit 'voice for the agent era' positioning and the first time it has shipped frontier-class reasoning + voice in a single coordinated launchSource: Xiaomi product site (mimo.xiaomi.com), Gizmochina, Artificial Analysis listing · 2026-04-22
- 1M-context surcharge removed at launch on Pro -- Xiaomi explicitly priced parity between short-context and full-context calls. Watch whether they reintroduce a tier later as adoption scales; the no-surcharge stance is unusual at this scaleSource: Xiaomi launch comms, Gizmochina · 2026-04
- ASR is open-source; TTS is API-only. If you need a fully self-hostable voice pipeline you can use ASR locally + a different TTS (ElevenLabs, Cohere, Murf) on top, or wait for Xiaomi to potentially open the TTS weights laterSource: Xiaomi announcement · 2026-04
- PRC content filtering applies on the reasoning/chat surfaces. Same regulated-topic pattern as Hy3, DeepSeek, Qwen, Kimi, GLMSource: Pattern across Chinese frontier APIs · 2026-04
Best for
Teams building voice-first agentic products that need a coordinated reasoning + TTS + ASR stack from a single vendor. Also Chinese-market builders and developers who need strong multimodal (vision + audio) inputs in one API call without stitching three providers together. MiMo-V2.6-Pro's single price across 1M context suits long-document agentic workloads, and its MIT open weights suit teams that want to self-host a flagship-class omnimodal model.
Not for
English-first creative writing (Claude / GPT-5.5 still lead), regulated geographies that block Chinese AI APIs, or teams whose only voice need is English TTS (ElevenLabs is more mature). Also not the right fit if you need independent benchmark verification, since most V2.6 numbers are Xiaomi's own.
Our Verdict
MiMo is Xiaomi treating voice as a first-class agentic surface, not an after-the-fact integration. Shipping V2.5 Pro + Multimodal + TTS + open-source ASR together -- with native vision and audio reasoning baked into the flagship and the 1M-context surcharge removed -- is the most coordinated voice-stack launch from a Chinese frontier vendor in 2026. MiMo-V2.6 (September 2026) adds MIT open weights for Pro and Flash at unchanged API prices, and Xiaomi's own tables put V2.6-Pro near Claude Opus 5 on several agent benchmarks while trailing on Terminal Bench 4.0 and exploit tasks. Treat MiMo as a serious option for voice-pipeline builds, multimodal Chinese-language workloads, low-cost agentic API use, self-hosted open-weight deployments, and dialect-strong ASR. For text-only English-first work, Claude / GPT / Gemini still lead and DeepSeek is still the cheapest text-first frontier alternative.
Sources
- Xiaomi MiMo blog: Introducing MiMo-V2.6 series (2026-09-22) (accessed 2026-10-29)
- Hugging Face: XiaomiMiMo/MiMo-V2.6-Pro-RL (accessed 2026-10-29)
- Hugging Face: XiaomiMiMo/MiMo-V2.6-Flash-RL (accessed 2026-10-29)
- Hugging Face: XiaomiMiMo/MiMo-V2.6-Pro-MOPD (accessed 2026-10-29)
- Hugging Face: XiaomiMiMo models (accessed 2026-10-29)
- Xiaomi MiMo docs: MiMo-V2 series deprecation (2026-06-30) (accessed 2026-07-04)
- Xiaomi: MiMo-V2.5-Pro product page (accessed 2026-04-25)
- Gizmochina: Xiaomi introduces MiMo-V2.5 TTS and ASR full voice pipeline (accessed 2026-04-25)
- Artificial Analysis: MiMo-V2.5-Pro listing (accessed 2026-04-25)
- The Asian Mirror: Xiaomi MiMo V2.5 voice AI launch (accessed 2026-04-25)
Explore more MiMo (Xiaomi) rankings
Deeper leaderboards, benchmarks, task-specific tier lists, and status/pricing pages for MiMo (Xiaomi).
The Tier List Tuesday
Weekly newsletter: tier movers, new entrants, and the VS of the week. Built from our daily AI-tool sweeps. No spam, unsubscribe anytime.
Alternatives to MiMo (Xiaomi)
Claude (Anthropic)
**Claude Duet opened as a research preview on 2026-10-21: a proactive agent app on iOS, Android and claude.ai, built around Claude Opus 5.5, that sends a morning brief from your connected apps, keeps working on a cloud computer while you are away and asks before it sends, pays or deletes anything. It is included in Pro and Max in the US with a waitlist, on the same usage limits as the rest of Claude.** **Oct 23: chat and Cowork became one Claude for Enterprise (GA). Oct 22: the full Claude Platform went GA on Google Cloud as an Anthropic-managed offering billed through Google Cloud. Oct 26: Anthropic filed an S-1 for an IPO on Nasdaq under ANT.** **Claude Haiku 5.5 shipped on 2026-10-07 at $0.10 / $0.50 per 1M tokens for prompts up to 100K ($0.50 / $2.50 above) -- 90% below Haiku 4.5 on most requests -- with adjustable effort for the first time on a Haiku, and the same day Anthropic halved Sonnet 5.5 cache reads to $0.10 and added monthly API credits for Max ($100 or $200) and Team (up to $500 pooled).** **Anthropic expanded the Cyber Verification Program on 2026-10-06 into three access tiers (Defense, Red Team, Specialized) that unlock reduced cyber safeguards on Opus 5.5, Sonnet 5.5, Mythos 5.1 and future models for verified security teams -- closing the 'coming weeks' note Opus 5.5 shipped with; generally available Claude keeps its conservative cyber blocks.** **Two Claude 5.5 models in six days: Claude Opus 5.5 (2026-09-22) and Claude Sonnet 5.5 (2026-09-28).** Opus 5.5 performs at Fable 5.1 level on most work and costs 40% less than Opus 5 to run -- $4/$20 per 1M (down from $5/$25) with cache reads cut to $0.20 (from $0.50), fast mode at $8/$40 up to 2.5x speed, output 30%+ faster, and five-hour usage limits raised on Pro, Max, Team and seat-based Enterprise with a saveable rate-limit reset. Sonnet 5.5 keeps Sonnet 5's $2/$10 but needs far fewer tokens (up to 30% cheaper per task), runs 30%+ faster and jumps Terminal-Bench 4.0 from 10.3% to 70.6%. Both are 1M context, Jun 2026 cutoff, on the Claude Platform, Bedrock, Vertex and Foundry; Haiku 5.5 followed on 2026-10-07. Opus 5.5 launches with Fable-class cyber and bio safeguards (most cyber tasks fall back to Opus 4.8) and cannot run with thinking off.
Claude Mythos 5.1
**From 2026-10-06 Mythos 5.1 is included in every tier of Anthropic's expanded Cyber Verification Program, and existing Project Glasswing members move into its Specialized Access tier without reapproval -- with the first disclosed Glasswing numbers: at least 129,000 verified vulnerabilities found April-July 2026, more than 33,000 rated critical or high.** Anthropic's trusted-access frontier model. **Mythos 5.1 launched 2026-09-01** alongside Fable 5.1, and Anthropic now states outright that they are **the same model with different safeguards** -- the 5.1-cycle gap is 60.9% vs 55.8% on Terminal-Bench 4.0, which Anthropic attributes to safeguard interventions rather than capability and expects to shrink. Originally launched June 9, 2026 alongside Claude Fable 5. Suspended June 12 by a US export-control order, then PARTIALLY RESTORED July 1, 2026 (US government lifted controls June 30): Mythos 5 is back for a set of US organizations with government approval, while Anthropic works to re-expand the broader Glasswing program. Public Fable 5 returned globally the same day. Gated to Project Glasswing orgs + select biology researchers.
Gemini (Google)
**Oct 29: the Gemini app can now generate, preview and edit documents, PDFs, slide decks, spreadsheets, Markdown and code right inside the chat for all users (PDF on web and mobile, the other formats on web only). Oct 30: the Computer Use tool went generally available in the Gemini API across the Interactions and GenerateContent APIs.** **Gemini 4 Carbon was announced on 2026-10-28 as Google's strongest model, but it is not generally available yet: access starts this week for a subset of cyber defenders through the Fairwind Program, and Google says developers, businesses and consumers get it later. When it opens it carries Argon's intro price of $2 / $10 per 1M tokens, rising to $4 / $20 after the intro period. Google reports it #1 on the Artificial Analysis Intelligence Index v5.0 at 59.** **Gemini 4 Argon went generally available on 2026-10-13: in the Gemini API (paid tier only), AI Studio, Gemini for business and GitHub Copilot, in the Gemini app and AI Mode in Search for Google AI Pro and Ultra subscribers, and in Google Antigravity for Ultra. The API rate card prices it at $2 input / $10 output per 1M tokens with $0.10 implicit-cache reads through November 30, 2026, doubling to $4 / $20 / $0.20 on December 1; there is no free tier and no explicit caching.** Google calls it its flagship reasoning model, with a 1M-token unified input-output limit, #1 on the Vals Index, 87% on the ERQA embodied-reasoning benchmark and 0.7% attack success on Gray Swan prompt injection, alongside the 9/30 numbers (77.9% DeepSWE v1.1, 51.3% AutomationBench, 91.7% LVBench, 68% CWE-bench v1) -- all vendor-reported. Earlier in October: Gemini 3.7 Flash and 3.5 Flash were retired from AI Studio (10/08) with API calls auto-routed to 3.8 Flash and 3.6 Flash, Nano Banana 2.1 went GA (10/06) and SynthID Detector opened to everyone in English (10/07). Late September in the Gemini app: **skills replace Gems** (9/30; slash-command instructions with reference files, Gems removed from November 2026 for personal accounts), **Guided Vision in Gemini Live** for blind and low-vision users on Android (10/01), and the 9/23-9/24 wave still stands: Gemini 3.8 Flash TTS and Flash-Lite TTS with voice design at $0.50/$9.00 and $0.50/$6.00 per 1M (doubling 2027-01-01), Gemini 3.8 Live with Live Avatar in Gemini Enterprise, free 1080p Omni 1.1 video in Google Vids. Googlebook was scheduled to reach shelves 10/04 (shipment not yet vendor-confirmed).
Grok
**Nov 2: the grok-imagine-image-quality API slug is retired; requests to it, and to grok-imagine-image-pro, which already redirected there, are now answered by grok-imagine-image-2.0 at low quality, which xAI prices $0.01 per image below the old model at every resolution ($0.04 / $0.05 / $0.06 at 1K / 1.5K / 2K).** **The X Subscription (2026-10-12) folds Grok, Grok Bot, Cursor and X Premium into one plan: Free, Plus $8, Premium $30 (what SuperGrok, Cursor Pro and X Premium+ cost $90 for combined), Super $100 and Ultra $200 a month, with one weekly-refreshing usage pool shared across Grok and Cursor; the old SuperGrok, Cursor and X Premium plans are closed to new sign-ups.** **Team Bots launched 2026-09-28 in public beta on Teams and Enterprise plans** -- shared Grok Bots built around a role, with team context (files, instructions, skills), plugins (Salesforce, Notion, GitHub), credentials for APIs without a plugin, and memories, where the bot is shared but each person's conversations and memories stay private; each Team Bot gets its own Slack handle. SpaceXAI runs one per major sales account, and its 9/22 case study says the merged SpaceXAI + Cursor support desk absorbed a 175% ticket increase on Grok Bot with no new hires at $0.20-$0.30 per resolution. **Grok 4.7 (9/21)** remains the flagship at $2/$6 per 1M with chart-only benchmarks; Grok Voice Transcribe 2.0 and the corrected $15/1M-character TTS price stand.
Muse Spark + Muse Wonder (Meta)
Meta Superintelligence Labs' model family, now four lines deep: **Muse Wonder** (2026-10-12, frontier, $4/$20 per 1M on the Meta Model API; Wonder Full only in Muse Code subscriptions), **Muse Spark 1.3** (2026-09-02, $1.25/$4.25, or $0.10/$0.20 in Contributor mode if Meta may train on your content), **Muse Spark Open** (2026-10-15, 816B/42B-active open weights, 1M context, runs on two Mac Studios with M5 Ultra) and **Muse Glimmer** (2026-08-10, 30B, Apache 2.0, built for local agents). Consumers get it through **Muse**, Meta's personal-agent app (2026-09-08): free with limits, Power $20/mo, Maximum $100/mo
GPT-Rosalind (OpenAI)
**Since 2026-10-15 US university labs can apply for the Rosalind Academic Plan: up to 30 ChatGPT Business Standard seats free for up to 12 months, with GPT-Rosalind and Rosalind Workbench behind trusted-access approval.** OpenAI's first domain-specific model -- life sciences, drug discovery, translational medicine. Launched 2026-04-16 as a Trusted Access research preview. Launch partners: Amgen, Moderna, Allen Institute, Thermo Fisher. Paired with a Life Sciences Codex plugin (50+ scientific tool integrations)
GPT-5.6-Cyber / GPT-5.4-Cyber (OpenAI)
**GPT-5.4-Cyber reached its API removal date on 2026-10-01: the gpt-5.4-cyber model page now returns Not found while gpt-5.6-cyber stays live (checked 2026-10-02; OpenAI's deprecations doc had not yet moved the row to Past Deprecations). Its replacement is `gpt-5.6-cyber`**, an alias for OpenAI's most advanced purpose-trained cyber models, gated behind the Daybreak program and priced at $12.50/$75 per 1M (first published price for any OpenAI cyber model). Original page: OpenAI's defensive-cybersecurity variant of GPT-5.4, launched 2026-04-16. Lowered refusal boundary for security-research tasks and native binary reverse-engineering. Access gated via Trusted Access for Cyber (TAC) program -- thousands of verified defenders, hundreds of teams, no public pricing. On **2026-08-17 OpenAI published its first dedicated post on the Hugging Face incident**, conceding it 'underestimated the real-world cyber capabilities of our AI models' and confirming it now releases cyber capabilities only to trusted defenders. **On 2026-09-01 OpenAI confirmed GPT-6 Astra meets the Critical cyber threshold -- the first model it has ever designated at that level -- and on 2026-09-03 committed $1B to Daybreak for Frontline Defenders**
Microsoft MAI-Thinking-1
Microsoft's first in-house reasoning model -- launched 2026-06-02 at Build as the flagship of seven new MAI models. 35B-active / ~1T-total sparse Mixture-of-Experts, 256K context. AIME 2025 97.0%, matches leading models on SWE-Bench Pro, and beat Claude Sonnet 4.6 in human-preference testing. Available on Microsoft Foundry + OpenRouter / Fireworks / Baseten
Tencent Hy3 / Hy4 preview (formerly Hunyuan)
Tencent Hy (formerly Hunyuan) open-weight MoE family -- Hy4 preview (2026-08-28) is the new flagship at 770B total and 49B active with a 1M-token context, Apache 2.0, at $0.834 in / $2.501 out per 1M tokens on Tencent Cloud TokenHub. Hy3 (GA 2026-07-06; 295B total / 21B active, 256K context, Apache 2.0) remains the budget pick at $0.132 / $0.528. Integrated into Yuanbao, WorkBuddy, CodeBuddy, and ima