NEW 1.6 TRILLION Open Source AI vs Claude & GLM | LongCat-2.0 🫣

Product reviewed

LongChat 2.0 / “LongCat-2.0” — an open-weight AI model described as 1.6T parameters. The creator team allegedly released an 8-bit quantized version (with the full-weight repository reportedly taken down). The video tests it via a free cloud service and compares quality/performance against multiple models (e.g., Claude, GLM, Kimiko, Gemini, DeepSeek) as well as against local quantized runs.

---

Key features / specs mentioned

• Scale: “massive open-weight” model, 1.6 trillion parameters
• Release status: full-weight repo reportedly taken down, but an 8-bit quantized version is available
• Storage / RAM requirements (as stated by narrator):
  • 8-bit: ~2 TB storage
  • 4-bit: ~1 TB memory
• Unique architectural note: claims n-gram embeddings are used in the architecture (noted as a differentiator from prior LongChat/LongChat Flash)
• License: MIT license (emphasized as very open and permissive)
• Service: tested through longchat.chat (free cloud service)

---

Benchmark / comparison claims (other models)

• Benchmarks claimed against:
  • Gemini 3.1 Pro (zipped)
  • Claude Opus (old/new references)
• Reported placement:
  • “Better than Gemini definitely”
  • In SWE: lands just above Opus ~4.6 and below ~4.7
• Not directly tested vs the narrator’s “top two” (in the narrator’s opinion):
  • Kimiko 2.7s
  • GLM 5.2

---

Tested user experience (prompts + results)

1) “Super Mario”-style 2D platform game generation

LongCat 2.0 outcome (cloud)

• Generates a platform-adventure game with Mario-like characters, coins, and bad guys
• Noted issues:
  • Inverted controls (A goes right, D goes left)
  • Aimless gameplay / no end flag
• Visual quality: described as beautiful and good-looking

Comparison

• Gemini: generates a “horrible” version (in this test)
• Opus (Claude, prior test): LongCat here is suggested as potentially as good or better, depending on whether “thinking” is enabled/disabled

Local comparison

• Best local generation: GLM 5.2 (4.8-bit) on a Mac Studio 512
  • ~51,700 tokens
  • ~4,100 seconds (~1 hour 8 minutes) to generate
  • Quality described as fantastic, including an end purpose and good combat interaction
• LongCat wasn’t highlighted as best; gameplay logic is described as mixed

---

2) Photorealistic real-time human face rendering (code)

LongCat 2.0 outcome (cloud)

• Errors about variable definitions, then attempts a fix
• The attempted fix is described as incorrect (narrator notes the issue was order: aspect defined after resize, so it must be above it)
• Final result: “worst generation / completely broken,” despite some “ambition”
• Some positives:
  • Skin tone works
  • Includes features like aging and expression controls
  • Expression “doesn’t work” (as described)

Comparison

• Narrator suggests configuration/settings could be involved (temperature, missing full version, API configuration, etc.)
• Local best face generation: Kimiko 2.7 (4.2-bit) using a clustered Mac Studio + MacBook Pro (128GB RAM)
  • ~7,500 tokens
  • ~15 tokens/second
  • Output praised as one of the best human head generations seen
• Conclusion: LongCat “maybe isn’t the best for coding,” though it could excel at specific tasks (discussed further below)

---

3) “AGI/lyrics logic” tests (“Rise Up”)

Prompt variant:

“If you rise up like the day and you rise up unafraid, how many days will you be rising up for?”

LongCat 2.0

• Gives a lyric-aligned answer via reasoning, but is described as inconsistent
  • First: answers “a thousand” (narrator says it’s a lyric reference problem)
  • Then: narrator cites an English vs non-English mismatch
    • Asked in English → “as many as it takes” (narrator disputes)

Other models

• GLM 5.2: reaches “every single one” style response
• GLM 4.7: “a thousand times” with a lyric reference (“I’ll do it a thousand times again”)
• DeepSeek (next thinking mode): “1,000 days”, treated as correct by narrator
• Claude/Sonnet 5: knows it’s “Rise Up,” but hallucinates “over and over…” phrasing not present in that line
• Kimiko 2.6: hallucinates (“sun grants… then one more”)

Result (as framed by narrator)

• “Winners: DeepSeek and GLM 4.7 (with lyric grounding)”
• Local LongCat consensus is described as wrong again

---

4) “Pronunciation from subtitles/lyrics” test (Rise Up: “each other”)

Prompt: ask how to pronounce a specific lyric.

Narrator’s point

• These models are trained on text and may not truly know sung pronunciation.
• Subtitles won’t contain audio/phonetic details, limiting performance.

LongCat 2.0 (cloud/local)

• Returns wrong pronunciation (narrator’s view: “each other” vs intended “each other’s/each other”, depending on what the narrator claims is correct)
• Repeats similarly incorrect answers even with max/high thinking locally

Other models

• GLM 5.2: provides phonetic guidance; narrator says it fails the lyrics/audio context anyway
• DeepSeek & Claude:
  • DeepSeek “thinks” longer (32–34 seconds) and responds tied to what the user “pronounced”
  • Narrator calls this “answered well but mission failure”
  • Claude max thinking becomes confused and also fails

Bottom line

• “No one actually got that test right,” likely due to lack of multimodal/audio/phonetic grounding

---

5) “Where am I?” UK lyric test (“Iron Maiden” / band-song)

Prompt asks where the singer is if they really like it and it’s wicked.

LongCat 2.0

• Says Beastie Boys, provides full line, but doesn’t identify the location
• Narrator calls this a failure

Other models

• GLM 5.2: thinks the user is in Boston / New England (wrong)
• DeepSeek: says it’s a skit; closer, but still wrong
• Claude Sonic 5: guesses dance floor (only “technically” aligned with the narrator’s reasoning)
• Kimiko 2.7: also guesses New England/Boston
• Local LongCat / local GLM 5.2:
  • GLM 5.2 loops into lyric possibilities and gets lost; even guesses Cardi B

Final takeaway (narrator)

• Models fail this “contextual lyric/location” task across the board; AGI/general intelligence isn’t there yet

---

Pros mentioned

• Open weights + MIT license (strong accessibility/open ecosystem point)
• Strong general performance vs Gemini
• Near/close SWE benchmark placement vs Claude Opus
• Good creative generation potential (Mario-esque look; described as “beautiful”)
• Some improvements with long thinking modes (can reason better in some logic/lyric situations)

---

Cons mentioned

• Local usability is difficult due to extremely large footprint (as stated):
  • 8-bit ~2TB / 4-bit ~1TB
• Coding correctness issues, including incorrect fixes (example: face-render variable-definition fix)
• Mixed generation reliability:
  • Some outputs look good but lack functional gameplay (e.g., no end condition)
• Hallucinations / lyric grounding issues:
  • Mistakes about lyric meaning and wrong references (e.g., Beastie Boys vs expected song)
• “AGI” test failures, especially:
  • Sung pronunciation
  • UK lyric/location inference

---

Overall verdict (concise)

LongCat 2.0 is promising but not a clear “best model” yet. It shows potential—especially in some creative generations—and is notable for its openness (MIT license) and innovation. However, in the video’s hands-on tests—particularly coding correctness and lyric/audio pronunciation-related tasks—it often underperforms GLM 5.2 / Kimiko 2.7 / DeepSeek / Claude.

The narrator implies the specific 1.6T / 8-bit quantized release may not be the best version to bet on locally, but suggests that future Flash/smaller versions could offer a better balance of quality vs usability.

Recommendation: Try it via cloud first; wait for smaller Flash versions or newer iterations, especially if your priority is coding accuracy or grounded reasoning.

---

Unique points mentioned about the product (deduplicated)

1. 1.6T parameter open-weight model (LongChat/LongCat 2.0)
2. Full-weight repo reportedly taken down; released 8-bit quantized weights
3. Quantization needs (per narrator): 8-bit ~2TB, 4-bit ~1TB
4. Architectural claim: n-gram embeddings
5. MIT license emphasized as very open
6. Tested via free cloud endpoint: longchat.chat
7. Claimed benchmarks: “better than Gemini” and close to Claude Opus in SWE (between ~4.6 and ~4.7)
8. “Mario” game: visually good, but inverted controls and missing end condition
9. Face rendering: errors and an incorrect fix; output described as “completely broken”
10. Lyrics/AGI tests: inconsistent grounding; hallucination and wrong references
11. Pronunciation test fails across models due to likely lack of multimodal/audio grounding
12. UK lyric location test fails (wrong artist/location); local runs may loop and drift

---

Speakers / views

• Single main speaker/narrator drives all tests.
• Comparative references are model-to-model results (no separate human speakers), with performance attributed to:
  • Claude (Opus/Sonnet)
  • GLM (5.2, 4.7)
  • Kimiko (2.7, 2.6)
  • Gemini (3.1 Pro / Opus references)
  • DeepSeek (various thinking modes)
  • “Sonic” (alongside Claude in narration)