Open-Source ChatGPT Alternative (2026, Free)

UIs added 2026-10-19: Open WebUI, LibreChat, Jan. Compare Open WebUI vs ChatGPT. Identification page: ChatGPT.

Listed models: LFM3-1.7B (Liquid AI text 1.73B / 256K, 635 likes, added 2026-11-01; not Omni; compare vs LFM3-Omni-2B), Mage-VL-Cowork (Microsoft Apache-2.0 ~4.7B hybrid CUA VLM, 93 likes, added 2026-11-01; not Iris-3B T2I), Granite Switch 4.1 (IBM Apache-2.0 8B+30B compose, 55 likes on 8B, added 2026-11-01), Clover-35B-A3B-Preview (DiffuSpace Apache-2.0 diffusion MoE, 35B/~3B, 128K, 176 likes, added 2026-10-31; compare vs Flint), LFM3-Omni-2B (Liquid AI 2.17B omni, text+speech out, 227 likes, added 2026-10-31; not LFM3-1.7B text), Mistral Large 4 (ungated ~1T/52B multimodal MoE, Ecosystem Open-Weight; MaaS needs a separate Mistral commercial license; 1,301 likes, added 2026-10-30; compare vs Muse Spark Open), LFM3 (Liquid AI ~2B edge family, 160 likes on LFM3-1.7B, added 2026-10-28; not LFM2.5-24B-A2B; compare vs LFM2.5-24B-A2B), Spark-X2.6-4B (XHToken Apache-2.0 ~4B on-device agent, 514 likes, added 2026-10-26; not Meta Muse Spark Open; compare vs LFM2.5-24B-A2B), Nanbeige4.3-33B (Nanbeige / BOSS Zhipin Apache-2.0 33B text teacher, 512K, 365 likes, added 2026-10-26; compare vs Qwen3.8-27B), Qwen3.8-27B (Qwen Apache-2.0 27B native VLM, 18,393 likes, identification page for that vs pair), DeepSeek-V4.1-Flash (1017 checkpoint), GLM-5.3-Flash (Z.ai MIT 320B/18B multimodal MoE, 2,941 likes, added 2026-10-25; compare vs Flash-1017), MiMo-V2.6-Flash (Xiaomi Flash-RL, card MIT tag / no LICENSE blob, 309B/15B, 582 likes, added 2026-10-25), MiniMax M3.1 Flash (MiniMax Community License 428B/23B multimodal MoE, 259 likes, weights last modified 2026-10-23; compare vs Flash-1017), Raon-1-mini (KRAFTON Apache-2.0 28.8B/4.3B preview, 148 likes, added 2026-10-24), Ornith-1.5-122B (MIT 122B MoE, added 2026-10-22), Flint (Apache-2.0 39.8B/2.5B local coding MoE, 533 likes, added 2026-10-23), Sage-A1.5 (Apache-2.0 35B science/agent VLM, 223 likes, added 2026-10-23), Carbon-X (Apache-2.0 14.8B DNA+text chat, 74 likes, added 2026-10-23; not a general ChatGPT stand-in and not clinically validated), Reflection Beam, Kimi K3, Hy-Memory, Ling-3.1-flash, MiniCPM-V 4.7, Muse Spark Open, Nemotron 3.5 Super VL, and Nemotron 3.5 Bolt. Other open LLMs also live on closed LLM APIs (today: Kolibri-1, LFM2.5-24B-A2B; LFM3 and Nemotron 3.5 Bolt are listed on both hubs). Image generators are on Midjourney-class alternatives. Compare Mistral Large 4 vs Muse Spark Open, LFM3 vs LFM2.5-24B-A2B, Spark-X2.6-4B vs LFM2.5-24B-A2B, Nanbeige4.3-33B vs Qwen3.8-27B, GLM-5.3-Flash vs DeepSeek-V4.1-Flash, MiniMax M3.1 Flash vs DeepSeek-V4.1-Flash, Flint vs Ornith-1.5-122B, Sage-A1.5 vs Ornith-1.5-122B, Ornith-1.5-122B vs DeepSeek-V4.1-Flash, Beam vs Muse Spark Open and DeepSeek-V4.1-Flash vs Kimi K3. Coding agents: Claude Code alternatives. Workspace clones: Notion alternatives (AppFlowy). Workspace AI: Notion AI alternatives. Pixel editors: PhotoCraft vs GIMP.

Sourced snapshot

Field	DeepSeek-V4.1-Flash (2026-10-13)	MiniCPM-V 4.7 (2026-10-15)	Muse Spark Open (2026-10-16)	Nemotron 3.5 Super VL (2026-10-17)	Nemotron 3.5 Bolt (2026-10-17)
Host	huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash (not gated)	huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B (not gated)	huggingface.co/meta-models/Muse-Spark-Open-816B-A42B (gated: manual)	huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 (not gated)	huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16 (not gated)
HF likes / downloads	4,462 / 1,272,612	631 / 4,157	264 / 402	142 / 40	50 / —
License	MIT	Apache-2.0 (card + GitHub LICENSE)	Muse Model License (SPDX other; limited commercial)	OpenMDW-1.1 (commercial-friendly)	OpenMDW-1.1 (commercial-friendly)
Size	552B backbone / 8B prefill · 16B decode	~35.2B BF16 (35B-A3B MoE); 1B variant ~1.55B	816B total / 42B active sparse MoE	120B total / 12B active hybrid MoE	74B total / 9B active LatentMoE
Job	Images + text in, text out	Image/video/document + text in, text out (on-device VLM)	Text + image + audio + video in, text out	Text + image + video in, text out	Text + image + video in, text out (local-agent VLM)
Context	1M tokens	—	1M tokens	1M tokens	256K tokens
Hardware	Vendor figures omitted	iPhone 17 Pro Max claim (vendor)	≥700 GB VRAM (NVFP4, vendor)	2× H100 or 1× GB200 (NVFP4, vendor)	24GB VRAM (IQ2 variant, vendor)
Quality vs ChatGPT	Not independently tested	Not independently tested	Not independently tested	Not independently tested	Not independently tested

October 19 additions

Field	DeepSeek-V4.1-Flash-1017 (2026-10-19)	Reflection Beam (2026-10-19)	Kimi K3 (2026-10-19)	Hy-Memory (2026-10-19)
Host	huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017 (not gated)	huggingface.co/reflection/Beam-1.0-501B-A23B (not gated)	huggingface.co/moonshotai/Kimi-K3 (not gated)	huggingface.co/tencent/Hy-Memory-30B-A3B (not gated)
HF likes / downloads	1,333 / 10,233	148 / 1,591	11,761 / 1,030,720	188 / 1,187
License	MIT	Apache-2.0	Kimi K3 License (SPDX other)	Apache-2.0
Size	552B backbone / 8B prefill · 16B decode	501.3B / 23.4B sparse MoE	2.8T / 104B LatentMoE	30B / ~3B writer MoE
Job	Images + text in, text out	Text in, text out (coding/agents)	Image-text-to-text	Memory writer for external stores
Context	1M tokens	1M tokens	1M tokens	Not independently verified
Quality vs ChatGPT	Not independently tested	Not independently tested	Not independently tested	Not independently tested

Arena.ai Code Arena WebDev (open-source filter, 18 October 2026 snapshot) lists DeepSeek-V4.1-Flash-20261017 (Max) at open-source #1 (score 1670) and Kimi K3 (Max) at #2 (1655). That is a third-party leaderboard, not a forkalt ranking and not a ChatGPT column.

Why DeepSeek-V4.1-Flash is on a ChatGPT hub

“Open source ChatGPT alternative” is a durable query. The new-term we could source on 2026-10-13 from a public, ungated card is DeepSeek-V4.1-Flash: MIT, 552B backbone, 8B/16B active, 1M context, 4,462 likes, created 2026-09-10. It natively takes images and text and emits text. That is the same job family as a hosted chat product, not a proof that it substitutes for ChatGPT’s applications, memory, or browsing.

ChatGPT is OpenAI’s hosted chat product. We use the ChatGPT mark only to identify the search. We did not copy OpenAI pricing or model names into a quality table.

Local versus hosted

The MIT weights are on Hugging Face. The card also links chat.deepseek.com, which is a hosted UI. If your requirement is “run weights on hardware I control,” start from the Hugging Face repo and the inference/ notes on the card — and budget for a 552B-backbone MoE. If your requirement is “a free chat box,” that hosted UI is a different artifact and is not what this hub ranks.

There is no GitHub repository deepseek-ai/DeepSeek-V4.1-Flash (404 on 2026-10-13). Encoding helpers are at deepseek-ai/deepseek-recipe.

What we left out

Other open chat checkpoints (including models already identified elsewhere on forkalt) will join this list when we attach them explicitly. Kolibri-1 stays on the generic closed-API hub so German/English MoE facts are not rewritten as a ChatGPT claim. MiniCPM-V 4.7 is attached here as an on-device multimodal option, not as a ChatGPT quality ranking. Muse Spark Open is attached as a gated Meta multimodal MoE identification row — datacenter-class VRAM, not a ChatGPT substitute.

Self-reported GPQA / Terminal-Bench / DeepSWE numbers on the DeepSeek card are quoted on the model page and labeled self-reported. They are not a ChatGPT comparison.

How to choose

• Need Mistral’s ungated ~1T/52B multimodal MoE (text+image, 1M, Ecosystem Open-Weight; MaaS extra license): Mistral Large 4. Compare Mistral Large 4 vs Muse Spark Open.
• Need Liquid’s text 1.73B / 256K card (not Omni, not LFM2.5): LFM3-1.7B. Compare LFM3-1.7B vs LFM3-Omni-2B. Family map: LFM3. Dedicated omni card: LFM3-Omni-2B. Compare family vs the older MoE: LFM3 vs LFM2.5-24B-A2B.
• Need Microsoft’s Apache-2.0 ~4.7B hybrid computer-use VLM (GUI + bash): Mage-VL-Cowork. Not Iris-3B.
• Need IBM’s Apache-2.0 8B+30B Switch (chat + decision aLoRA + embedded TurboCTC): Granite Switch 4.1. Speech-only family: Granite Speech 5.0.
• Need DiffuSpace’s Apache-2.0 diffusion 35B-A3B (128K, Qwen3.6 base): Clover-35B-A3B-Preview. Compare Clover-35B-A3B-Preview vs Flint.
• Need Apache-2.0 ~4B on-device text (Ollama/llama.cpp named; not Meta Muse Spark): Spark-X2.6-4B. Compare Spark-X2.6-4B vs LFM2.5-24B-A2B.
• Need a 33B Apache-2.0 text teacher (512K, agent/coding card): Nanbeige4.3-33B. Compare Nanbeige4.3-33B vs Qwen3.8-27B.
• Need Qwen’s 27B native VLM (262K, Apache-2.0 LICENSE file): Qwen3.8-27B.
• Need Z.ai’s MIT 320B/18B multimodal MoE: GLM-5.3-Flash. Compare GLM-5.3-Flash vs DeepSeek-V4.1-Flash.
• Need Xiaomi’s 309B/15B Flash-RL card (MIT tag; no LICENSE blob): MiMo-V2.6-Flash.
• Need MiniMax’s 428B/23B multimodal MoE and can accept the Community License (attribution + $20M revenue trigger): MiniMax M3.1 Flash. Compare MiniMax M3.1 Flash vs DeepSeek-V4.1-Flash.
• Need a compact Apache-2.0 preview MoE (28.8B/4.3B, Korean+agent card): Raon-1-mini.
• Need Apache-2.0 39.8B/2.5B local coding MoE with GGUF/Ollama: Flint. Compare Flint vs Ornith-1.5-122B.
• Need Apache-2.0 35B science/agent VLM (256K, orchestration/verification): Sage-A1.5. Compare Sage-A1.5 vs Ornith-1.5-122B.
• Need MIT 122B MoE coding/agent weights and a vLLM/SGLang recipe: Ornith-1.5-122B. Compare Ornith-1.5-122B vs DeepSeek-V4.1-Flash.
• Need MIT multimodal weights with a 1M context figure and the 1017 checkpoint: DeepSeek-V4.1-Flash.
• Need MIT 560B/25B hybrid MoE (card 1M context; config.json 262,144): Ling-3.1-flash.
• Need ungated Apache-2.0 text MoE for coding/agents (501B/23B, Mirror CLI): Reflection Beam. Compare Beam vs Muse Spark Open.
• Need 2.8T-class open multimodal weights and can accept the Kimi K3 License: Kimi K3. Compare DeepSeek-V4.1-Flash vs Kimi K3.
• Need an agent memory writer (30B-A3B) plus SDK/MemHub, not a chat model: Hy-Memory.
• Need Apache-2.0 on-device VLM with a 35B-A3B MoE plus a 1B sibling: MiniCPM-V 4.7. iPhone 13+ tok/s / 3 GB is a vendor claim.
• Need gated Meta 816B/42B multimodal MoE (1M context; datacenter-class VRAM): Muse Spark Open. Blog quality adjectives stay vendor claims.
• Need ungated NVIDIA datacenter VLM (120B/12B, 1M context, OpenMDW-1.1 commercial-friendly): Nemotron 3.5 Super VL. Compare with Nemotron 3.5 Super VL vs Muse Spark Open.
• Need local-agent VLM that fits in 24GB VRAM (74B/9B, 256K context, Ollama-ready): Nemotron 3.5 Bolt. Vendor RTX 84 tok/s claim.
• Need NVIDIA speech / file ASR, not a chat VLM: Nemotron 3.5 Transcribe on Otter.ai alternatives. Do not mix that id with Bolt or Super VL.
• Need the Aleph Alpha German/English MoE already sourced on this site: Kolibri-1.
• Need a self-hosted chat UI (branding clause): Open WebUI — Open WebUI vs ChatGPT.
• Need a MIT self-hosted UI: LibreChat.
• Need an offline desktop app: Jan.
• Need DeepSeek’s official inference engine for the 1017 checkpoint: DeepInfer (Apache-2.0, 13,104 stars, 2026-10-21) — DeepInfer vs vLLM.
• Need a general Apache-2.0 LLM server: vLLM (94,459 stars, v0.32.0, 2026-10-21).
• Need embeddings rather than chat: OpenAI embeddings alternatives.

What “open” does not mean

DeepSeek-V4.1-Flash weights are MIT and not gated. That is a stronger “take the files home” story than LTX-2.5 or SAM 3. It is still a 552B-backbone MoE with 1M context in the README. “Free” on this hub means the card’s price is not a SaaS SKU; it does not mean a laptop GPU will load the full model. We have not published a quantization recipe.

The card’s comparison table names Opus-5.0, GPT-5.6 Sol, K3, GLM-5.3, and DeepSeek’s own V4-Pro / V4-Flash. ChatGPT as a product is not a column. Treating any of those rows as “beats ChatGPT” would be editorial malpractice. Self-reported GPQA Diamond 90.9 and Terminal-Bench 2.1 90.6 stay on the model page with selfReported: true.

Prompt encoding is not a default Jinja template. If you wire this checkpoint into a generic OpenAI-compatible proxy without encoding/ or deepseek-recipe (381 GitHub stars on 2026-10-13), you are off the official path. Hosted chat.deepseek.com is convenient and not the MIT weights.

See also closed LLM APIs, classification APIs, Cursor alternatives (Headroom compression layer, Orca orchestrator, DeepSeek Harness MIT runtime — added 2026-10-25), Perplexity alternatives, and AI Chat & Agents. Serving stacks added 2026-10-21: DeepInfer, vLLM, DeepInfer vs vLLM. Community chat (not an LLM UI): Discord alternatives.